A learning algorithm for rate selection in real-time

Computer Networks 126 (2017) 114–124

Contents lists available at ScienceDirect

Computer Networks journal homepage: www.elsevier.com/locate/comnet

A learning algorithm for rate selection in real-time wireless LANs Michele Luvisotto a,∗, Federico Tramarin b, Stefano Vitturi b a b

Department of Information Engineering, University of Padova, Via Gradenigo 6/B, 35131, Padova, Italy National Research Council of Italy, CNR–IEIIT, Via Gradenigo 6/B, 35131, Padova, Italy

a r t i c l e

i n f o

Article history: Received 15 February 2017 Revised 14 June 2017 Accepted 3 July 2017 Available online 8 July 2017 Keywords: Real–time WLANs Learning algorithm SNR estimation Rate selection

a b s t r a c t To achieve a real–time behavior in wireless communication systems, the multi–rate support (MRS) provided by the IEEE 802.11 Wireless LAN standard may reveal particularly advantageous. Unfortunately, the most widespread rate adaptation algorithms designed for general purpose applications proved to be unsuitable for the challenging real–time scenario. This has led to the definition of purposely designed algorithms such as RSIN, a rate adaptation technique based on the SNR measurement, which showed very good performance in terms of timeliness and reliability. The goal of this paper is to propose an improvement of RSIN that extends its applicability to a wider range of applications. To this aim, we introduce RSIN–E, an enhanced version of RSIN based on an estimation of the SNR obtained through a learning algorithm. In detail, this paper first provides an exhaustive description of both the proposed learning algorithm and the relevant estimation procedure. Then it presents an extensive performance assessment of RSIN–E carried out via both experimental sessions and simulations. The obtained results confirm the effectiveness of the proposed technique and highlight that its performance figures are comparable with those of RSIN, and significantly better than those of Minstrel, a widespread rate adaptation algorithm adopted by most general purpose applications. © 2017 Elsevier B.V. All rights reserved.

1. Introduction Real–time communication systems are requested to provide performance figures that are quite different from those of general purpose communications [1]. Significant examples, to this regard, are represented by networks deployed in factory automation systems as well as in motion control applications. Indeed, the typical industrial traffic is characterized by the cyclic transfer of small data amounts, as well as by the transmission of short acyclic messages, generated by alarms, with very tight deadlines. Similarly, networks adopted by motion control applications may have cycle times (i.e. the duration of the intervals in which a controller carries out a complete polling of a set of sensors/actuators) as low as some hundreds of microseconds, with very low jitters, while an alarm message should be notified within a deadline of the same order of magnitude. Furthermore, the appearance of real–time multimedia industrial applications, such as video surveillance and remote tracking, has led to the introduction of another type of traffic, characterized by the exchange of greater data amounts that, nevertheless, maintains tight timing requirements.

∗

Corresponding author. E-mail addresses: [email protected] (M. Luvisotto), [email protected] (F. Tramarin), [email protected] (S. Vitturi). http://dx.doi.org/10.1016/j.comnet.2017.07.002 1389-1286/© 2017 Elsevier B.V. All rights reserved.

federico.

To deal with such kinds of traffic, different communication systems have been deployed in the years, specifically, fieldbuses [2], real–time Ethernet networks [3] and, more recently, real–time (industrial) wireless networks [4]. Among the different wireless technologies that can be adopted, the IEEE 802.11 Wireless LAN (WLAN) [5] represents an important opportunity, since it can be adequately tailored to achieve performance figures almost comparable with those of the wired real–time counterparts [6]. A feature of IEEE 802.11 that significantly contributes to its adoption in this context is multi–rate support (MRS). Basically, MRS allows to dynamically change the transmission rate to adapt to the possible fluctuations of the channel status. Thus, since low rates use more robust modulations, they can be selected when the channel is in a bad state (i.e., when the signal–to–noise–ratio, SNR, is low), whereas higher rates can be used for better channel states. This kind of strategy is implemented by suitable rate adaptation (RA) algorithms that, however, are not explicitly specified by the IEEE 802.11 standard. The design of effective RA algorithms is hence an interesting and widely targeted research topic [7]. The most popular ones nowadays adopted are arguably the Automatic Rate Fallback (ARF) [8] and Minstrel [9]. These algorithms aim at maximizing the throughput, irrespective of the number of transmission attempts carried out at the Medium Access Control (MAC) layer [10]. Such a behavior, unfortunately, results unsuitable for the majority of real–time applications, since the random backoff waiting times

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

introduced between subsequent transmission attempts clearly impair the communication timeliness. As a consequence, some RA algorithms specifically tailored for real–time industrial communication have been designed [11]. In this respect, in [12] we proposed a technique, called RSIN (Rate Selection for Industrial Networks), that showed effective performance in this specific application framework. As it will be better illustrated in the next sections, RSIN relies on the knowledge, by a station that is about to send a frame, of the SNR perceived by the intended frame receiver. Based on such an information, as well as on the deadline imposed to frame delivery (which is a typical constraint of real–time communication), and on the relation between Packet Error Rate (PER) and SNR, RSIN implements an optimization algorithm that provides the station with both the maximum number of MAC layer attempts and the list of rates to be used during the transmission procedure. This strategy maximizes the probability that the frame is correctly delivered within the deadline. The knowledge of the SNR value is clearly of prominent importance for the appropriate operation of RSIN. To this regard, in [12], where a cyclic polling protocol between a master device and some slaves was considered, the SNR was actually measured by each slave upon receiving the poll–request frame and then encapsulated in the payload of the poll–response frame sent back to the master. However, it has to be noted that this behavior limits the range of applications in which RSIN can be profitably exploited, since in several cases the SNR value might not be available to a station willing to deliver a frame. For example, unidirectional communication applications do not require response frames and, hence, a sender can not rely on this strategy to know the SNR perceived by the receiver. Also, it may happen that the SNR measurement is impaired by unpredictable factors, leading to suboptimal rate selection. These observations represent, actually, the starting point for this paper. To extend the applicability of RSIN to a more widespread range of applications, in this work we introduce a new, enhanced, version of RSIN called RSIN–E, that leverages a learning approach to estimate the SNR. In this framework, the previous transmission history is exploited by a station as an input to the learning algorithm to provide an accurate estimation of the channel status. With respect to [12], the main contributions of this paper can be summarized as follows • it introduces the SNR estimation procedure adopted by RSIN– E; • it provides the results of an extensive set of measurements carried out via both experimental sessions as well as simulations, in different operational contexts, to assess the performance of RSIN–E; • it compares the obtained results with those of both RSIN and Minstrel, showing definitely that RSIN–E represents an interesting opportunity for real–time communication based on WLANs. In detail, the paper is organized as follows. Section 2 gives a description of RSIN. Section 3 introduces the strategy adopted to estimate the SNR and discusses its main parameters. Section 4 describes the adopted experimental setup and provides the results of a practical assessment carried out on RSIN–E, RSIN and Minstrel. Section 5 presents the results of some realistic simulations executed to investigate the behavior of RSIN–E and to compare it with that of RSIN, on a more complex network configuration. Finally, Section 6 concludes the paper and outlines some directions for future research activities. 2. The RSIN algorithm The RSIN algorithm is executed at the MAC layer of each station in the network before a new packet transmission takes place

115

and provides the station with both the number of transmission attempts and the rate at which each attempt has to be carried out, chosen among the ones available at the underlying physical (PHY) layer. In order to work effectively, this technique requires two fundamental assumptions to be verified: 1. The sending station must know the value s of the SNR1 . 2. The sending station must know the relation F between the SNR and the PER for each available rate and for a frame of the same size (L) of the one it has to transmit. If the above assumptions are verified, and indicating with R =

{R1 , . . . , RR } the set of available rates, the station can compute

the residual transmission error probability for a frame of length L bytes carried out at the MAC layer with the  after N attempts  rates r (1 ) , . . . , r (N ) ∈ RN . Let us indicate this residual error prob-





ability with P L, s, N, r (1 ) , . . . , r (N ) . Moreover, the station can compute the maximum time necessary to deliver the packet (taking into account the backoff procedure), which is  IEEE 802.11 random  indicated as D L, N, r (1 ) , . . . , r (N ) . The two variables introduced are used by a constrained optimization procedure which constitutes the core of RSIN: the algorithm selects the number of transmission attempts and the corresponding sequence of rates that minimize the packet error probability, while ensuring that the maximum delivery time does not exceed the deadline associated to the packet, D. More formally, this corresponds to the solution of the following minimization problem

min N≤Nmax

,r (i ) ∈R



P L, s, N, r (1 ) , . . . , r (N )



(1)

subject to the constraint





D L, N, r (1 ) , . . . , r (N ) ≤ D

(2)

where Nmax is the maximum number of transmission attempts allowed at the MAC layer. The importance of the constraints expressed by both Eqs. (1) and(2) in the framework of real–time industrial communication can be explained considering that, in case of cyclic traffic, the deadline of a frame typically reflects its transmission period, whereas for acyclic traffic the deadline is associated to the latency of the alarm carried by the transmitted packet [13]. 3. SNR estimation for RSIN On the foundation provided by RSIN, in the next part of the paper we present an improved rate adaptation scheme named RSIN– E. In particular, we provide a mathematical framework that allows the transmitter to achieve an accurate estimation of the SNR based on the analysis of the previous transmissions. Consequently, the performance of RSIN–E does not rely on the ability of measuring the SNR anymore. 3.1. General framework Fig. 1 proposes a schematic representation of RSIN–E. In the lower part, it clearly highlights the original SNR estimation block, whose description will be detailed in this section. The legacy RSIN optimization phase is found in the upper part of the figure. As described in Section 2, it outputs the transmission rates r (1 ) , . . . , r (N ) for the subsequent frame delivery, based on the input map F and the SNR level. Such a map has to be made available in advance. 1 We assume in the following that s lies in a finite set S of possible values, which has cardinality  .

116

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

which indicate the frequency with which the different rates were selected during the kth observation period. With these premises, the task of the SNR estimation algorithm is to find the SNR value which better explains the observed error probabilities defined in Eq. (4), based on the a priori knowledge represented by the PER vs. SNR map F. To solve the aforementioned problem, we define the following cost function

∀s ∈ S, E (s, k ) =

R 

2

wi (k )[F (s, Ri ) − Pi (k )]

(7)

i=1

Fig. 1. Schematic representation of the RSIN–E algorithm.

The estimation block leverages on the fact that the final correct reception (or discarding) of a data frame is acknowledged by an ACK (or by its missed reception). Consequently, the transmitter gains a precise knowledge of the performance of the previous transmissions and can compute the error probabilities P1 , . . . , PR for any of the adopted rates. Combining this information with the PER vs. SNR map F, the needed SNR estimate sˆ can be obtained, through the procedure detailed in the following. 3.2. SNR estimation algorithm A first and significant assumption is that the given estimation procedure is implemented as a periodic process. That is, a new SNR estimate is provided every update period, whose duration is Tu . Consequently, at the end of the kth update period, based on the observed performance, a transmitting station has to refresh its estimation of the SNR for the subsequent (k + 1 )th period, namely sˆ(k + 1 ). Clearly, the update period must be selected carefully: if it is too short, there will not be enough data to use for the estimation; conversely, if it is too long, the wireless channel status might have changed significantly during the observation period so that the obtained estimation reveals not adequate for the forthcoming transmissions. Nevertheless, it is possible to provide reasonable bounds for Tu . Indeed, since the typical real–time industrial traffic is to a large extent cyclic, even if with different periods, the lowest transmission period Tp represents a lower bound to Tu , which ensures that at least one packet transmission attempt has occurred within the update period. On the other hand, the channel coherence time Tc constitutes an upper bound for Tu , ensuring that the channel remains stationary in that fraction of time [14]. A general rule of thumb for the choice of the update period is hence

Tp ≤ Tu ≤ Tc

(3)

After the value of Tu has been selected, for each update period k, let Si (k) be the number of observed successful transmission attempts when the rate Ri was employed, and let Ai (k) be the total number of transmission attempts performed in k with the same rate Ri , with Si (k ) ≤ Ai (k ), i = 1, . . . , R. The observed error probabilities during the kth period for each different rate can be derived as:

Ai ( k ) − Si ( k ) , Ai ( k )

Pi (k ) =

i = 1, . . . , R

(4)

Furthermore, the total number of successes and attempts during the k–th period can also be computed as:

S (k ) =

R 

Si (k ),

i=1

A (k ) =

R 

Ai ( k )

(5)

i=1

We may also define the following weights:

wi ( k ) =

Ai ( k ) , A (k )

i = 1, . . . , R

(6)

which represents, for each SNR value s, the square difference between the expected error probabilities according to the map F and the observed error probabilities during the kth period. For each rate Ri , the weighting factor wi , defined in Eq. (6), is used to give more importance to the rates which have been selected more frequently during the observation period. The SNR estimation for the kth period is hence defined as

sˆ(k + 1 ) = arg min E (s, k )

(8)

s∈S

that is, the SNR value which minimizes Eq. (7) for a given k. Although Eq. (8) is actually the most obvious solution to the presented SNR estimation problem, the typical behavior of industrial real–time traffic and wireless channel non–idealities demand for further refinements to obtain adequate performance. Indeed, during an update period of length Tu only few channel observations may be available, often relevant to a single frame transmission, i.e. Tu ࣃ Tp . Clearly, this avoids an effective tracking of wireless channel variations, impairing the accuracy of the SNR estimation, which in turn results not stable over time. For this reason, the SNR estimation procedure has also to take into account the past history of transmission attempts, not limiting to the last observation period. Consequently, we modify the estimation problem of Eq. (8), which we set up as a regularized optimization problem [15]. Specifically, we define a penalty function H, designed in such a way that the estimated SNR values leading to high error probabilities are penalized. To this aim, let A˜ (s, k ) be the total number of MAC layer transmission attempts up to the k-th period, where we indicate with s the (estimated) SNR value at which those attempts have been carried out. Analogously, let S˜(s, k ) be the corresponding number of successful transmissions. These two functions, differently from the ones included in Eq. (4), are not related to the adopted transmission rates but only depend on the values of the estimated SNR adopted in the previous update periods. Therefore, given that in the kth update period the estimated SNR value is sˆ(k ), A˜ (s, k ) is updated as follows



∀s ∈ S, A˜ (s, k ) =

A˜ (s, k − 1 ) + A(k ) if s = sˆ(k ) A˜ (s, k − 1 ) otherwise

(9)

whereas the update of S˜(s, k ) is performed analogously. Exploiting these quantities, the penalty function for the update period k can be defined as:



H (s, k ) =

0 A˜ (s,k )−S˜(s,k ) A˜ (s,k )

˜ S˜(s,k ) if A˜ (s, k ) = 0 or A(s,kA˜ )− < Pth s,k

(

otherwise

)

(10)

In particular, Eq. (10) shows that H (s, k ) generally corresponds to the error probability observed in the whole transmission history up to period k, when s was selected as the estimated SNR. However, if either a particular SNR s was never estimated in the past, i.e. A˜ (s, k ) = 0, or the observed error probability is smaller than a threshold Pth , whose significance will be discussed in detail in the next subsection, then no penalization takes place.

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

117

Algorithm 1 Selection of penalty coefficient λ.

Fig. 2. Tuning penalty coefficient in RSIN with SNR estimation: initial training phase and subsequent running phase.

Introducing now a penalty coefficient λ ∈ [0, 1], it is possible to weigh the contributions of the cost function E (s, k ) and the penalty function H (s, k ). If λ is large, then the penalty function (and hence the history of the network) is weighted more; conversely, if λ is small, the cost function (and hence the results of the last observation) period takes more importance. To this aim, the following new objective function is defined

E (s, k ) = (1 − λ )E (s, k ) + λH (s, k )

(11)

Consequently, the final SNR estimate for each update period Tu can be obtained as the solution of the following regularized estimation problem

sˆ(k + 1 ) = arg min E (s, k ) s∈S

(12)

3.3. Tuning of the parameters The first meaningful parameter to be tuned is the update period Tu , discussed in Section 3.1 and, in general, regulated by Eq. (3). In the applications discussed in this paper the update period is chosen quite low (i.e., Tu ࣃ Tp ), as a prompt response to channel variations is needed. Moreover, the estimation of the channel coherence time Tc is generally a tricky task, highly dependent on the surrounding environment. Hence, a clear upper bound for Tu might not be available. The selection of the probability threshold Pth also plays an important role in the value of the penalty function defined in Eq. (10). Pth is introduced to avoid penalizing those SNR values that, when estimated, have led to a relatively low error probability. If Pth is high, this would happen for too many SNR values, thus decreasing the weight of the penalty function H (s, k ). A small value for this parameter is hence generally selected, for instance in this paper we set Pth = 0.1. Furthermore, the penalty coefficient λ strongly influences the performance of the SNR estimation and, consequently, of the RSIN–E algorithm. The role of λ is to balance the minimization of the cost function in Eq. (7) with that of the penalty function in Eq. (10). An optimal value for λ cannot be derived a priori and depends on several factors, such as packet size, transmission period, nodes positions, channel impairments, etc. Nonetheless, the typical application scenarios for real–time networks are often characterized by well–defined traffic profiles and mobility patterns. This is the case, for instance, of industrial real–time communication applications, where traffic follows a cyclic pattern and nodes usually have a very limited mobility compared to general purpose wireless networks (e.g., cellular networks). Consequently, in the context of industrial networks, a strategy for the experimental tuning of λ can be devised, as roughly represented in Fig. 2 and, more detailed, in Algorithm 1. After the network is deployed, a training phase is initiated, during which the nodes placement and traffic flows reproduce exactly those of the subsequent running phase. During this training phase the penalty coefficient is varied among different values picked from a set 

1: procedure PenaltyCoeffSelection(, Aλ , λ , Tλ ) λopt ← 0 2: 3: J opt ← 0 4: J curr ← ∞ 5: while true do 6: t←0 curr −J opt > λ then 7: if J Jopt 8: for each l ∈  do λ←l 9: 10: Perform Aλ transmission attempts 11: Al ← Aλ 12: Sl ← number of successful attempts 13: J (l ) ← AlA−Sl l 14: end for opt λ ← arg min J (l ) 15:

 Initializations

 Network running forever  Reset time  Training phase

l∈

16: 17: 18: 19: 20: 21: 22: 23:

λ ← λopt J opt ← J (λopt ) end if  Running phase Run the network until t = Tλ AR ← number of total attempts during the running phase SR ← number of total successes during the running phase R J curr ← ARA−S R end while

24: end procedure

and, for each value, relevant information on the transmission outcomes are collected. At the end of this phase, a metric of specific interest, called J, is computed based on the collected information. For example, in Algorithm 1, the collected information are the number of transmission attempts Al and successes Sl for each value l ∈  and the metric J is computed as the total percentage of failed transmissions. Finally, the optimal value of the penalty coefficient λopt is selected as the one that minimizes the metric J and the corresponding value of the metric is stored in Jopt . The network then goes on with a running phase, where the value of the penalty coefficient is fixed to λopt . Clearly, this value represents the best choice for the current network configuration and wireless channel status, provided that the training phase is long enough to collect a meaningful amount of data. Nevertheless, in the long term, the operating conditions can change (for example due to the sudden disturbance by an external interferer). As a result, an adaptive tuning of λ is hence needed to ensure that the selected value always yields the best performance under the actual operating conditions. To this aim, during the running phase, the nodes keep collecting information on the transmission outcomes and, after a time Tλ , the metric J is computed again and stored in Jcurr . At this point, as reported in both Fig. 2 and Algorithm 1, each node checks if the relative error between the current value of the metric Jcurr and the previously computed optimal value exceeds a threshold λ

J curr − J opt > λ J opt

(13)

If the threshold is exceeded, a new training phase is started, which leads to the definition of new values for λopt and Jopt . Conversely, the network continues the running phase, maintaining the penalty coefficient fixed. Clearly, the sensitivity and delay with which RSIN–E detects and reacts to changes in the operating conditions depend on the choice of the parameters Tλ and λ . 3.4. Computational complexity Since the SNR estimation phase has to be performed online, it is crucial that its computational burden is limited. To this regard, the following considerations can be made. First, at each period Tu , the update of both Ai (k) and Si (k) requires only R operations each. The computation of probabilities Pi (k)

118

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

and weighting factors wi (k) also requires R operations each. In addition, the update of A˜ (s, k ), S˜(s, k ) and the penalty function H (s, k ) requires  operations each, one for each SNR value in S. Finally, the computation of the cost function E (s, k ) requires R ·  total operations. Therefore, the complexity of the proposed estimation algorithm is O(R ·  ), being R( + 4 ) + 3 the total number of operations that have to be carried out at each update period. For instance, in a system where R = 8 transmission rates are available, and considering 30 dB as the range of variability for the SNR values (with a granularity of 1 dB), then a total of 362 elementary operations are required. Considering the whole procedure involved in RSIN–E with reference to Fig. 1, its complexity is actually dominated by both the discussed estimation algorithm and the optimization step carried out by the legacy RSIN. As already pointed out in [12], the computational burden associated to RSIN may actually result demanding. However, it is worth observing that the whole algorithm can be reduced, under some specific hypotheses, to a simple search within a look–up table. Indeed, if the frame length L is constant (as it often happens in real–time industrial networks) and the measured PER vs. SNR map F does not change significantly over time, the optimization phase of RSIN that leads to the generation of the sequence of rates can be carried out offline for each value of the SNR s ∈ S and stored in the memory of the station. Then, at run time, after the SNR value has been estimated, it is sufficient to search for the corresponding solution. Stemming from the above considerations, it may be concluded that RSIN–E can be effectively adopted even by simple industrial devices such as sensors/actuators. Finally, it is worth mentioning that RSIN–E has been implemented following the aforementioned guidelines in the devices of the experimental set–up described in the next section. 4. Performance assessment in a prototype network In this section the performance figures of RSIN–E will be experimentally evaluated in different scenarios and compared with those of both the legacy RSIN and Minstrel. 4.1. Experimental setup In order to validate the approach proposed in this paper, the

RSIN–E algorithm has been implemented in a prototype network based on commercial off–the–shelf Wireless Network Interface Cards (WNICs). The implementation is based on the IEEE 802.11 stack provided by the Linux kernel and, specifically, on the mac80211 module, following the approach detailed in [9,12]. The optimal rate selection of RSIN–E is executed before any packet transmission procedure, in order to provide the WNIC driver with the optimal rates and number of attempts, whereas the SNR estimation is performed periodically with period Tu . Since both these procedures require the knowledge of the PER vs. SNR map F, a preliminary measurement campaign has been conducted to characterize precisely the relation between SNR and PER for different packet lengths, as explained in [16]. The modified mac80211 modules have been loaded in a set of desktop workstations (Dell Optiplex models 745, 755 and 960) with Ubuntu 14.10 operating system based on Linux kernel version 3.16.4. Each workstation is equipped with a TP-LINK WNIC (models TL-WN851ND and TL-WN881ND), based on the AR9287 chip, which is handled by the open–source ath9k driver. Both WNICs are compliant with the IEEE 802.11n standard and support Multiple Input Multiple Output (MIMO) operations up to a 2 × 2 architecture. Specifically, the Space–Time Block Coding (STBC) strategy is enabled in order to increase the communication reliability [17]. Moreover, 40 MHz channels are used in the 2.4 GHz band and a regular guard interval is employed, following the guidelines for

Fig. 3. Prototype IEEE 802.11n network used for the experimental validation.

industrial IEEE 802.11n networks presented in [16]. This particular choice of parameters made available R = 8 different Modulation and Coding Schemes (MCSs) with transmission rates ranging from 13.5 to 135 Mbit/s. For this particular PHY configuration, the preliminary experimental campaign allowed to observe that the meaningful range of SNR values goes from 7 to 36 dB. Indeed, for SNRs lower than 7 dB a PER of 1 is observed for any MCS, while an SNR of 36 dB is high enough to guarantee successful transmissions for all MCSs. Consequently, the set S used for the SNR estimation contained  = 30 SNR values, ranging from 7 to 36 dB with 1 dB spacing. This quantization choice allowed a fair comparison with the legacy RSIN algorithm based on measured SNR, since both the received signal strength (RSS) reading and the noise floor in the adopted WNICs have a granularity of 1 dB. The desktop workstations have been arranged in a prototype IEEE 802.11n network composed of three nodes, as shown in Fig. 3. The network is configured in infrastructure mode and emulates a typical industrial application: a central controller, which acts as Access Point (AP), periodically sends request packets to distributed sensors/actuators, represented by WLAN stations (STAs), that send response packets when polled. This exchange of packets is realized through a purposely developed software application, installed in all nodes. The period with which the controller sends a request packet to each one of the nodes is fixed to Tp = 25 ms. To mimic typical real–time industrial applications, packets are exchanged at the data–link layer, avoiding network and transport protocols and hence, besides the MAC header and trailer, they only contain an application layer payload of length L. Two example values have been considered for the payload length: L = 50 bytes, which corresponds to a traditional industrial application involved with the exchange of very short commands and sensor readings, and L = 500 bytes, which may be instead representative of more advanced industrial multimedia applications (e.g., video surveillance) [18]. The prototype IEEE 802.11n network has been tested in a research laboratory, where a complete electromagnetic isolation was not feasible. However, the carrier frequency used by the WNICs has been carefully selected after monitoring the environment with a real–time spectrum analyzer, so that to avoid channels where other wireless networks were operating. Moreover, particular attention was dedicated to emulating typical wireless channel impairments found in industrial environments. To this aim, the channel models proposed by the IEEE 802.11 Task Group n (TGn) [19] have been taken into consideration and, specifically, model “F” was chosen, since it is the one that better represents the propagation characteristics of industrial buildings. In order to reproduce this channel behavior in the prototype network, an Agilent E4433B RF signal generator has been introduced in the setup, and configured to

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

119

Table 1 Parameters of the experimental setup. Description

Value

PHY/MAC standard MIMO configuration Channel bandwidth Modulation and Coding Schemes (MCSs) Transmission rates

IEEE 802.11n 2 × 2 STBC 40 MHz @ 2.4 GHz 0–7 13.5, 27, 40.5, 54, 81, 108, 121.5, 135 Mbit/s IEEE 802.11 TGn model “F” {50, 500} bytes 25 ms {0.5, 1, 2, 5} ms 10 {7, 8, . . . , 36} dB 25 ms 0.1 {0, 0.3, 0.6, 0.9} 10 0 0 250 s 0.1 0.6 (L = 50 bytes), 0.3 (L = 500 bytes)

Channel model Payload size (L) Polling period (Tp ) Packet deadline (D) Max. number of MAC–layer attempts (Nmax ) RSIN–E SNR range for estimation (S) RSIN–E update period (Tu ) RSIN–E probability threshold (Pth ) RSIN–E set of penalty coefficients () RSIN–E training phase attempts (Aλ ) RSIN–E penalty coefficient period (Tλ ) RSIN–E penalty coefficient threshold (λ ) RSIN–E optimal penalty coefficient (λopt )

generate AWGN–like noise centered on the operating frequency of the WNICs. The output power was varied to reproduce a quantized version of the channel gain behavior found in model “F”. The controlled noise produced by the RF generator has been injected on the channel through a directional antenna pointed towards all the employed WNICs, as depicted in Fig. 3. In the quantization of the channel model, three levels have been used: a low one, corresponding to an SNR level at which all available MCSs fail except the lowest one (which is however impaired); an intermediate one, which impairs only the highest MCSs; a high one, which corresponds to a high SNR level at which all MCSs can be used without producing any transmission error. The tuning of both the probability threshold for the SNR estimation and the update period has been carried out in agreement with the considerations reported in Section 3.3. Thus, for the RSIN–E algorithm, these two parameters have been set to the values Pth = 0.1 and Tu = 25 ms, respectively. Hence, the estimated SNR is updated at each polling cycle and the cost function in Eq. (7) is computed by taking into account only the outcome of the very last packet transmission. The choice of the penalty coefficient λ, has been carried out emulating the network training phase explained in Section 3.3. The metric to minimize, J, is represented by the percentage of failed pollings, as in Algorithm 1. Specifically, the set of candidate values for λ contained four values,  = {0, 0.3, 0.6, 0.9} and, during the training phase the outcomes of Aλ = 10 0 0 polling cycles have been recorded for each value in . The update parameters were set to Tλ = 250 s and λ = 0.1. It is worth observing that, since the laboratory environment is static and controlled, the condition in Eq. (13) was always observed and, hence, only one initial training phase was needed in our experiments. Specifically, the outcomes of this training phase were λopt = 0.6 for the case of short packets (L = 50 bytes) and λopt = 0.3 for longer ones (L = 500 bytes). This difference confirms the observation that the optimal value of λ depends on the traffic features, e.g. the payload size, and hence has to be carefully selected through a preliminary training phase. The meaning and value of all the parameters adopted in the experimental evaluation are reported in Table 1. 4.2. Qualitative assessment of the estimation performance A first assessment of the RSIN–E performance is shown in Fig. 4 in a qualitative way. The figure reports the estimated SNR and the corresponding chosen rates by the AP for 50 consecutive

Fig. 4. Qualitative assessment of the SNR estimation performance for L = 50 bytes and D = 0.5 ms: real, measured and estimated SNR (top) and corresponding MCS chosen for the first MAC–layer transmission attempt (bottom). In the MCS plot, for RSIN and RSIN–E a filled marker indicates a successful transmission, while an empty marker indicates a failed one.

polling cycles during the network running phase. In this assessment, short packets of L = 50 bytes have been exchanged (hence λ has been set to 0.6) and the deadline has been set to D = 0.5 ms. In detail, the top figure shows three different SNR patterns: the orange line with circular markers reports the real SNR value at the receiving STA, the dark blue line with diamond markers indicates the measured SNR sent to the AP (basing on which the legacy RSIN algorithm selects the optimal rates) and the light blue line with square markers reports the SNR estimated by RSIN–E. First, it is noticeable that the real SNR levels are approximately distributed around three main values, corresponding to the three channel quantization levels: 7 dB (bad channel), 20 dB (average channel) and 36 dB (good channel). Moreover, it is evident that the measured SNR follows exactly the real value with a delay of 1 transmission. Indeed, the measured SNR is retrieved from the response packet and, hence, corresponds to the value measured by the receiver during the previous packet transmission. As far as the SNR estimation is concerned, it may be observed that, although there is a certain difference between measured and estimated values, the general trends of the two curves are the same: when the measured SNR goes up (due to a high channel gain), also the estimated one increases, and the same happens when the SNR suddenly drops. The lower part of the figure shows the MCS selected by both versions of RSIN for the first packet transmission attempt, based on the SNR value. Specifically, the orange line reports the MCS choice for the real SNR, the dark blue line reports the MCS chosen by the legacy RSIN algorithm, and the light blue line reports the MCS chosen by RSIN–E. For RSIN and RSIN–E filled markers indicate a successful transmission, whereas empty markers indicate a failed one. Again, it can be observed that the trend of the RSIN–E algorithm follows that of the legacy RSIN. Moreover, it can be noticed that the former is generally more conservative, in the sense that most of times it chooses a rather lower MCS than the one chosen by RSIN. In particular, looking at the transmission attempts after #1785, it can be observed that RSIN–E evidently estimates a low SNR (and hence selects MCS 0), even if the real SNR is high. This is due to the fact that some previous transmission attempts (e.g., #1766, #1767 and #1774) failed despite the estimated SNR was good. This specific aspect of RSIN–E driven by the presence of the penalty function in Eq. (10), may increase the packet

120

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

Fig. 5. Histogram of MAC–layer transmission attempts for different rate adaptation algorithms and deadline values (L = 50 bytes).

Fig. 6. Histogram of the MCSs selected by the rate adaptation algorithms for different deadline values (L = 50 bytes).

transmission times even if the channel state is good but, on the other hand, ensures higher success probabilities when the channel state is average or bad, which may reveal a good strategy especially if a high reliability is mandatory [20]. 4.3. System performance with short packets In this section the performance figures of RSIN–E are compared with those of the legacy RSIN as well as with those of the widespread Minstrel algorithm tuned for real–time communications [21]. A first insight is provided by Fig. 5, which shows the histogram of the MAC–layer transmission attempts2 required for the delivery of a packet for the three different algorithms and two values of the RSIN deadline, D = 0.5 ms and D = 5 ms. Ideally, a good rate adaptation algorithm for real–time communication must perform the lowest possible number of transmission attempts, since each attempt adds significant delay and jitter due to the retransmission and backoff mechanisms of IEEE 802.11. It can be first observed that setting a lower deadline results, intuitively, in a lower number of transmission attempts: for both RSIN algorithms, when D = 0.5 ms no more than 2 attempts are ever required. Moreover, the distribution of transmission attempts for RSIN and RSIN-E is quite similar for a given deadline value, with only small percentage differences. Finally, when the Minstrel algorithm is employed, a non–negligible percentage of transmissions require a high number of attempts (6 or more), as this algorithm keeps on retransmitting a packet until it is received or the maximum number of attempts is reached. Fig. 6 shows the histogram of the MCSs that have been selected by the rate adaptation algorithm considering all transmission attempts at the MAC layer and, again, for D = 0.5 ms and D = 5 ms. Several observations can be drawn. First, when the deadline is short, the RSIN algorithm only selects two values: the lowest one (MCS 0), when the channel is bad, and the highest one (MCS 7), when the channel is good. The behavior of RSIN–E, instead, is different: MCS 0 is also the most frequently selected MCS, whereas MCS 7 is almost never selected, with MCS 2 and 4 being adopted instead. This is a reflection of the conservative behavior al2 The transmissions that required only one attempt are not reported here for the sake of clarity. The figure, hence, only reports the cases when MAC–layer retransmissions have been necessary.

Fig. 7. Percentage of failed pollings (lost or late packets) versus the deadline for different rate adaptation algorithms (L = 50 bytes).

ready spotted in Fig. 4. When the deadline is higher, the outcomes are slightly different but follow the same trend: RSIN selects either MCS 0, 2 or 7, whereas RSIN–E chooses mostly the first five MCSs. Finally, the Minstrel algorithm selects most frequently MCS 1 (which yields the best “average” performance), even if all the available MCSs are adopted a non–negligible amount of times, due to the sampling behavior of this technique [9]. The most important performance indicator for industrial real– time rate adaptation algorithms is arguably the percentage of failed pollings, reported in Fig. 7 for different deadline values, from D = 0.5 ms to D = 5 ms. The figure allows to distinguish between pollings failed because the packet (either request or response) was lost (lighter part of the bar) or because it arrived after the corresponding deadline (darker part of the bar). As a general trend, it can be observed that, for all algorithms, the higher the deadline, the better the performance, as it is intuitive. Moreover, in almost all cases, the performance figures of RSIN–E are slightly worse than those of RSIN but significantly better than those of Minstrel, the only exception being when the deadline is very high (D = 5 ms), since in this case Minstrel has enough time to perform many retransmissions. Specifically, looking at the percentage

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

121

Table 2 Service time statistics with L = 50 byte long request packets, for different rate adaptation algorithms and deadline values. Algorithm

Mean

Std. deviation

RSIN, D = 0.5 ms RSIN–E, D = 0.5 ms RSIN, D = 5 ms RSIN–E, D = 5 ms Minstrel

528.9 μs 529.6 μs 566.6 μs 558.3 μs 1061.1 μs

416.8 μs 309.1 μs 443.9 μs 429.1 μs 1770.4 μs

Fig. 9. Percentage of failed pollings (lost or late packets) versus the deadline for different rate adaptation algorithms (L = 500 bytes).

Fig. 8. ECDF of service time for L= 50 byte long request packets for different rate adaptation algorithms and deadline values.

of packets arrived after the deadline, it emerges that Minstrel violates significantly more deadlines than the RSIN and RSIN–E algorithms. This is linked to the fact that the latter are aware of the deadline and consider it in the rate selection, whereas the former does not. It is worth observing that the percentage of failed pollings results in general quite high, particularly if compared with that of general purpose applications typical of the everyday life. This is due to the specific test conditions. Indeed, the measurements have been carried out at the MAC layer with severe channel conditions and without any upper layer protocol. Consequently, error correction techniques, typical of transport/application layer protocols are not applied, with the inevitable impact on the packet error rate. Another important performance indicator is provided by Table 2, which shows the statistics of the service time (mean and standard deviation) for request packets (i.e., packets sent by the AP to the STAs), considering all rate adaptation algorithms and deadlines of D = 0.5 ms and D = 5 ms. As a first important observation, it is evident that the performance of Minstrel are extremely poor, both in mean and standard deviation, stemming from the fact that it is not designed for real–time communications. The high standard deviation, in particular, makes it very badly suited for real– time industrial applications. The performance of RSIN and RSINE, instead, are very similar in terms of mean and standard deviation. Furthermore, as can be seen, the deviation of the service time with RSIN–E results always slightly lower than that obtained with RSIN, due to the conservativeness of the former procedure as previously observed. The results of Table 2 are confirmed by Fig. 8, which shows the Empirical Cumulative Distribution Function (ECDF) of the service time for the different rate adaptation algorithms. The figure, however, allows obtaining some further important insights. Indeed, as can be seen, the ECDF of RSIN and RSIN–E when the deadline

is high (D = 5 ms) are practically overlapping and significantly outperform Minstrel in terms of packets delivered within 1 ms (more than 90%). Also, looking at the ECDF when the deadline is short (D = 0.5 ms), it can be observed that RSIN delivers a notable percentage of packets (more than 20%) in less than 400 μs, whereas RSIN–E never does it, since, due to its conservativeness, it almost never adopts the highest MCS. On the other hand, the two versions of RSIN are able to deliver a very high percentage of packets (more than 95%) within 1 ms. 4.4. System performance with long packets The growing complexity of networked control systems calls for different kinds of applications, where bigger amount of data must be exchanged with real–time constraints. For example, real– time industrial multimedia applications are concerned with the exchange of images and/or video frames that allow performing video–surveillance or real–time tracking of objects in an industrial setup [18]. In these cases, the typical payload length is much higher than that of the classical applications and may reach several hundreds of bytes. Thus, in a further experimental session, similar tests to those carried out in Section 4.3 have been performed with a payload length of L = 500 bytes for both request and response packets. In this case, according to the considerations made in Section 4.1, the penalty coefficient of the SNR estimation algorithm has been set to λ = 0.3. The minimum deadline employed for the RSIN optimization has also been increased from 0.5 to 1 ms. The percentage of failed pollings for the different rate adaptation algorithms is presented in Fig. 9. The biggest difference with the case of shorter packets in Fig. 7 is that the RSIN–E algorithm now outperforms the version of RSIN that relies on measured SNR for the shortest deadline values. Indeed, the higher conservativeness of RSIN-E is more effective when packets are longer, since the impact of retransmissions may lead to considerably longer transmission times, and likely results in missing the deadline, especially if such deadline is short. Moreover, the performance gap between RSIN and Minstrel is lower in this case, since Minstrel works better when the packet size is high, as highlighted in [21]. The statistics of the service time are reported in Table 3. As can be seen, both RSIN and RSIN–E have similar performance and are able to deliver a substantial reduction in mean and standard deviation of the service time with respect to Minstrel.

122

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

Table 3 Service time statistics with L = 500 byte long request packets, for different rate adaptation algorithms and deadline values. Algorithm

Mean

Std. deviation

RSIN, D = 1 ms RSIN–E, D = 1 ms RSIN, D = 5 ms RSIN–E, D = 5 ms Minstrel

737.8 μs 844.0 μs 804.7 μs 736.2 μs 1191.7 μs

551.7 μs 465.4 μs 694.7 μs 558.9 μs 1777.1 μs

Fig. 10. ECDF of service time for L= 500 byte long request packets for different rate adaptation algorithms and deadline values.

A final insight is given by Fig. 10, which reports the ECDF of the service time for request packets and presents a quite different situation with respect to Fig. 8. The Minstrel algorithm is again capable of providing very low service time values (around 500 μs), but a significant percentage of packets (more than 10%) takes more than 1.5 ms to be delivered. The legacy RSIN algorithm is also able to reach very low service time values (around 400 μs) with a good probability, whereas for RSIN–E the minimum values of the service time are around 500 μs (with D = 5 ms) and 700 μs (with D = 1 ms). However, RSIN–E guarantees an upper bound of the service time, since almost all packets are delivered within 1.5 ms (for both deadline values), contrarily to the performance of the legacy RSIN where the percentage of packets that takes more time is around 10%. 5. Simulations in a larger industrial network In typical real–world applications, more complex networks than those addressed in Section 4 may be deployed, where several stations are distributed in the environment and all communicating with a central controller. Unfortunately, reproducing such a kind of setups with desktop PCs in a research laboratory is challenging. Thus, in this section a simulative assessment is presented, aimed at evaluating the scalability of RSIN–E in larger networks. The performance assessment has been carried out with the ns3 network simulator [22], which includes a good implementation of the IEEE 802.11 MAC layer that has been validated against experimental testbeds [23]. To enhance the dependability of the simulations even further, the IEEE 802.11 PHY layer and the wireless channel of the legacy ns3 implementation have been modified according to the results obtained during the experimental campaigns. To this regard, as a first relevant upgrade, the experimentally

Table 4 Percentage of failed pollings for the two versions of the RSIN algorithm with different deadlines in a simulated industrial infrastructure network of M = 10 nodes (L = 50 bytes). Deadline

D = 0.5 ms D = 1 ms D = 2 ms

Failed pollings [%] RSIN

RSIN-E

19.32 17.79 15.43

13.65 12.41 7.64

measured PER vs. SNR curves mentioned in Section 4.1 have been inserted to determine whether a packet is successfully received or discarded at the PHY layer, basing on the SNR, the packet size and the transmission rate. Moreover, since the legacy wireless channel models of ns3 are mostly limited to large–scale fading effects such as path loss and shadowing, a more accurate model has been considered in this evaluation that, in addition to these impairments, introduces also an emulation of small–scale fading effects. Specifically, a quantized realization of TGn channel model “F” has been considered, mimicking the artificial noise introduced in the experimental evaluation of Section 4 through the RF generator. The ns3 platform upgraded with the aforementioned features has been used to simulate an IEEE 802.11n infrastructure network3 composed of one central controller (that acts as AP) and M = 10 attached nodes. A cyclic communication schedule is established, where the controller sequentially polls each node sending a request packet and receiving a response packet, both of size L bytes. The cycle period is set to Tcycle = 50 ms, with a slot assigned for the polling of each node set to 5 ms. A polling is considered as failed if the response packet does not arrive within the assigned slot. The nodes are randomly deployed on a circular area centered on the controller, with a minimum distance of 1 m and a maximum one of 3.5 m. It is worth highlighting that the results presented in this section have been averaged over 10 different random dispositions, to avoid dependence on a particular disposition of nodes. The simulative assessment has focused only on the comparison between the two versions of the RSIN algorithm, which have been both implemented in ns3. Different deadlines for the service time have been considered, with the maximum one limited to D = 2 ms, to allow for an exchange of two packets within the 5 ms slot. For what concerns the SNR estimation parameters, the update period Tu has been set equal to the period with which a single node is polled, i.e., the cycle time Tcycle = 50 ms, whereas the other parameters have been kept to the values used during the experimental evaluation, i.e., λopt = 0.6 and Pth = 0.1. A first set of results, concerned with the exchange of L = 50 byte long packets, is reported in Table 4, which shows the cumulative percentage of failed pollings. As can be seen, in the simulative assessment, RSIN–E always outperforms the legacy RSIN strategy. This result contrasts with what observed during the experimental evaluation and is mostly due to the fact that the conservativeness of RSIN-E allows a higher degree of robustness in an extremely controlled environment, such as the one simulated with ns3. In practice, RSIN–E tends to estimate a low SNR and hence selects the lowest rates (which correspond to the most robust modulation and coding schemes) more often than the legacy RSIN, as already noted in the experimental evaluation (see Fig. 6). While the conservativeness of RSIN-E reveals effective in terms of reliability, it may reveal detrimental for the timeliness of the polling procedure. Indeed, the legacy RSIN is able to immediately recognize good channel conditions, so that it reacts by selecting

3 The PHY and MAC configuration of the IEEE 802.11n WLAN corresponds to that of the experimental setup, with the parameter values reported in Table 1.

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124

123

Furthermore, the analysis carried out indicates some interesting future activities. In particular, the SNR estimation procedure needs to be experimentally validated in real environments, and hence additional tests (e.g., in the industrial scenario where tight performance figures are required) should be devised. Also, the actual implementation of RSIN–E on real devices (e.g. sensors/actuators) needs to be assessed. This will allow to definitely evaluate the performance improvements brought by the proposed technique. References

Fig. 11. Average polling time for the two versions of the RSIN algorithm with different deadlines in a simulated industrial infrastructure network of M = 10 nodes (L = 500 bytes).

the fastest transmission rates, while the estimation–based approach is more conservative and almost always prefers the slowest ones. To provide an example, Fig. 11 reports the average polling time of the two algorithms for different values of the deadline D for the case of L = 500 bytes packets. It can be observed that the legacy RSIN strategy is in general able to achieve better performance than RSIN–E. This is particularly evident when the deadline is low, that results in almost a half polling time with respect to RSIN–E. 6. Conclusions In this paper we presented RSIN–E, a rate selection algorithm specifically conceived for real–time applications of WLANs. Its key feature is the adoption of a learning algorithm for the estimation of the SNR. After giving a detailed description of the SNR estimation procedure, we provided the outcomes of a performance assessment of RSIN–E carried out experimentally, as well as by means of simulations. The obtained results are encouraging, since they highlight that the RSIN–E algorithm provides performance figures comparable with those of the legacy RSIN algorithm, which is based on the actual SNR measurement, and significantly better than those of Minstrel, in terms of reliability, timeliness and determinism. This makes RSIN–E a viable opportunity for the multirate support of WLANs deployed in demanding real–time scenarios. Moreover, it extends the applicability of the legacy RSIN algorithm to a wider range of applications, namely those in which a direct measurement of the SNR is not available or possible.

[1] S. Vitturi, P. Pedreiras, J. Proenza, T. Sauter, Guest editorial special section on communication in automation, IEEE Trans. Ind. Inf. 12 (5) (2016) 1817–1821. [2] J.P. Thomesse, Fieldbus technologies in industrial automation, Proc. IEEE 93 (6) (2005) 1073–1101. [3] T. Sauter, The three generations of field-level networks–evolution and compatibility issues, IEEE Trans. Ind. Electron. 57 (11) (2010) 3585–3595. [4] A. Willig, Recent and emerging topics in wireless industrial communications: a selection, IEEE Trans. Ind. Inf. 4 (2) (2008) 102–124. [5] IEEE Standard for Information technology–Telecommunications and information exchange between systems–Local and metropolitan area networks– Specific requirements. Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications, 2012. [6] Y.-H. Wei, Q. Leng, S. Han, A. Mok, W. Zhang, M. Tomizuka, RT-WiFi: real– time high-speed communication protocol for wireless cyber-physical control applications, in: Real-Time Systems Symposium (RTSS), 2013 IEEE 34th, 2013, pp. 140–149. [7] S. Lv, X. Wang, X. Zhou, On the rate adaptation for IEEE 802.11 wireless networks, Comput. Netw. 54 (17) (2010) 3173–3186. [8] J.-H. Yun, Performance analysis of IEEE 802.11 WLANs with rate adaptation in time-varying fading channels, Comput. Netw. 57 (5) (2013) 1153–1166. [9] W. Yin, P. Hu, J. Indulska, Rate control in the mac80211 framework: overview, evaluation and improvements, Comput. Netw. 81 (2015) 289–307. [10] L. Deek, E. Garcia-Villegas, E. Belding, S.-J. Lee, K. Almeroth, A practical framework for 802.11 MIMO rate adaptation, Comput. Netw. 83 (2015) 332–348. [11] S. Vitturi, L. Seno, F. Tramarin, M. Bertocco, On the rate adaptation techniques of IEEE 802.11 networks for industrial applications, IEEE Trans. Ind. Inf. 9 (1) (2013) 198–208. [12] F. Tramarin, S. Vitturi, M. Luvisotto, A dynamic rate selection algorithm for IEEE 802.11 industrial wireless LAN, IEEE Trans. Ind. Inf. (2016) 1. [13] G. Bernat, A. Burns, A. Llamosi, Weakly hard real–time systems, IEEE Trans. Comput. 50 (4) (2001) 308–321. [14] H. Jung, T.T. Kwon, K. Cho, Y. Choi, React: rate adaptation using coherence time in 802.11 WLANs, Comput. Commun. 34 (11) (2011) 1316–1327. [15] J. Friedman, T. Hastie, R. Tibshirani, The Elements of Statistical Learning, 1, Springer, Berlin, 2001. [16] F. Tramarin, S. Vitturi, M. Luvisotto, A. Zanella, On the use of IEEE 802.11n for industrial communications, IEEE Trans. Ind. Inf. 12 (5) (2016) 1877–1886. [17] S. Alamouti, A simple transmit diversity technique for wireless communications, IEEE J. Sel. Areas Commun. 16 (8) (1998) 1451–1458. [18] J.S. Blanes, J. Berenguer-Sebasti, V. Sempere-Paya, D.T. Ferrandis, 802.11N performance analysis for a real multimedia industrial application, Comput. Ind. 66 (0) (2015) 31–40. [19] V. Erceg, TGN channel models. Techinal Report 03/940r4, IEEE 802.11, 2004. [20] A. Willig, K. Matheus, A. Wolisz, Wireless technologies in industrial networks, Proc. IEEE 93 (6) (2005) 1130–1150. [21] F. Tramarin, S. Vitturi, M. Luvisotto, Improved rate adaptation strategies for real-time industrial IEEE 802.11n WLANs, in: Emerging Technologies Factory Automation (ETFA), 2015 IEEE 20th Conference on, 2015, pp. 1–8. [22] The Network Simulator - ns-3. [Online]. Available: http://www.nsnam.org/. [23] N. Baldo, M. Requena-Esteso, J. Núñez Martínez, M. Portolès-Comeras, J. NinGuerrero, P. Dini, J. Mangues-Bafalluy, Validation of the IEEE 802.11 MAC model in the Ns3 simulator using the EXTREME Testbed, in: Proceedings of the 3rd International ICST Conference on Simulation Tools and Techniques, 2010, doi:10.4108/ICST.SIMUTOOLS2010.8705.

124

M. Luvisotto et al. / Computer Networks 126 (2017) 114–124 Michele Luvisotto received the Masters degree in automation engineering from the University of Padova, Padova, Italy, in 2014. He is a PhD student with the Department of Information Engineering, University of Padova, Italy. He has been a visiting researcher at ABB Corporate Research Center in Vsters (Sweden) from March to September 2016. His research interests include wireless networks and real-time industrial communication.

Federico Tramarin is a senior Post-Doctoral researcher with the Institute of Electronics, Computer and Telecommunication Engineering of the National Research Council of Italy (IEIIT-CNR) since 2013.He got his PhD in Information Engineering in 2012 with the Electronic Measurement research group of the University of Padova, Italy, which he joined in 2008, after his master degree in Electronic Engineer.His research activities, started in the field of distributed measurements, are recently focused on industrial communication systems, performance characterization and protocol optimization for realtime (wired and wireless) networks.He is a member of the IEEE IES Technical Committee of Factory Automation, and of the IEEE P61158 Working Group.

Stefano Vitturi received the Laurea degree in electronics engineering from the University of Padova, Padova, Italy, in 1984.He has been a Senior Researcher with the Institute of Electronics and Computer and Telecommunications, National Research Council of Italy, Padova, since January 2002. From 1985 to 2001, he worked at the control and data acquisition system of RFX, a nuclear fusion experiment located in Padova, where he was the head of the Automation and Informatics group. His research interests include industrial automation systems and real-time industrial communication networks.