Abstract This paper presents a comprehensive perfor-mance study of closed-loop fast link adaptation (FLA) inthe context of IEEE 802.11n, spanning the physical (PHY)and medium-access control (MAC) layers. In particular, asemi-analytical model is derived for Basic and request tosend/clear to send (RTS/CTS) access schemes of the dis-tributed coordination function (DCF), that applies to both,open- and closed-loop strategies. Numerical results serveto demonstrate the accuracy of the proposed model and thesuperiority of FLA, in terms of MAC goodput, in compar-ison to open-loop policies. Realistic operating conditionssuch as outdated feedback information and the use of sta-tistical packet length distributions, issues not treated in pre-vious studies, have also been considered. Moreover, it isshown that incorporating a time-out mechanism in the FLAscheme, weighing down the influence of channel informa-tion as this becomes outdated, is a useful strategy to coun-teract its deleterious effects.
This work has been partially funded by MEC and FEDER throughproject COSMOS (TEC2008-02422), AM3DIO (TEC2011-25446)and Conselleria d’Educació, Cultura i Universitats del Govern de lesIlles Balears through a PhD grant.
G. Martorell (B) · F. Riera-Palou · G. FemeniasMobile Communications Group, Department of Mathematics andInformatics, University of the Balearic Islands, 07122 Majorca,Spaine-mail: [email protected]
Over the last decade, the IEEE 802.11 standard for wire-less local area networks (WLAN) has become the preva-lent technology for indoor wireless Internet access. Morerecently, and in response to the growing demands for highercapacity, the IEEE standards committee has published thefinal version of IEEE 802.11n [1] as a new amendmentof IEEE 802.11. Compared to previous specifications, thisnew norm allows much higher throughputs to be achievedwhile being able to fulfill more stringent quality of service(QoS) requirements, yet remaining fully backwards com-patible with previous versions of the standard. This amend-ment specifies enhancements to the IEEE 802.11 physicallayer (PHY) and the medium access control (MAC) sub-layer, most notably, the use of multiple-antenna technology(so-called MIMO) over the orthogonal frequency-divisionmultiplexing (OFDM) technique and frame aggregation, re-spectively. Additionally, it incorporates a feedback controlchannel from the receiver (Rx) to the transmitter (Tx) thatenables the implementation of closed-loop adaptive mecha-nisms.
Adaptation plays a crucial role in dealing with the timevarying nature of the wireless channel. Adaptive mecha-nisms allow the reconfiguration of system parameters in or-der to exploit the available instantaneous channel capacitywhile satisfying QoS constraints. One of the most widelyused reconfiguration techniques is adaptive modulation andcoding (AMC), which selects an appropriate modulation andcoding scheme (MCS) in response to changes in the environ-ment or system behavior. AMC algorithms can be broadlycategorized as closed- or open-loop, depending on whetheran explicit feedback channel between Rx and Tx is usedor not. Open-loop setups operate in a heuristic manner andtheir rate of adaptation tends to be slow with respect to chan-
nel changes, thus compromising the fulfillment of QoS con-straints. In contrast, closed-loop mechanisms track more ac-curately the channel behavior and are more reactive to rapidchannel variations.
Most IEEE 802.11-based systems employ the distributedcoordination function (DCF) with Basic Access scheme atthe MAC sublayer and adopt open-loop AMC policies suchas automatic rate fallback (ARF) [2] or one of its variants(e.g. CARA [3], SARA [4]). Owing to its simplicity, ARFis by far the most popular algorithm in use. However, theDCF with Basic Access scheme does not differentiate be-tween collisions and transmission failures caused by poorchannel conditions. Consequently, when the system expe-riences a high collision probability, ARF tends to use thelowest transmission rate even if the channel conditions arefavorable to use much higher transmission modes (see forexample, [3, 5–7]). Other adaptive strategies have been pro-posed to solve this issue, but they may require frame for-mat changes [8], modifications to the medium access tech-nique [3], or the use of channel quality indicators (e.g. sig-nal strength indicator) [7, 8] and, in fact, none of them hasachieved widespread use in current WLAN systems [9].
Alternatively, DCF can employ the request to send /clearto send (RTS/CTS) access scheme instead of the Basic Ac-cess scheme. The use of RTS/CTS shortens the collisionduration and allows to differentiate between collisions andtransmission failures at the expense of increased overhead.Consequently, RTS/CTS-based operation results in very dif-ferent system characteristics whose performance is worthstudying further.
Analytical models for DCF-based WLANs that do notmake use of AMC have been known since long ago [10–12],however, the inclusion of AMC in the theoretical frameworkhas only been recently addressed [13, 14]. These studies (see[13, 14]) have demonstrated that, in the context of IEEE802.11n, the use of closed-loop techniques such as fast linkadaptation (FLA) offers important benefits in terms of phys-ical layer throughput, yet how this improvement reflects onthe MAC goodput of a FLA-based system remains largelyunexplored. Based on the analytical model introduced byBianchi et al. in [10], and latter refined by Tinnirello et al.in [15] and Chen in [16], and complemented with simu-lation results obtained using the physical layer model de-scribed in [13, 14], this paper presents a semi-analyticalmodel that can be used to assess the goodput performanceat the MAC layer of both, open- and closed-loop adaptiveschemes targeting IEEE 802.11n. The proposed model ex-pands the one presented by the authors in [17] by model-ing the retry limits and the anomalous slot performance re-ported in [15], and the error-prone channel behavior intro-duced by Chen in [16]. Additionally, issues that may affectvery significantly the practical implementation of closed-loop strategies such as having to cope with delayed (possibly
outdated) feedback information and the possible utilizationof different packet lengths are also considered in this study.Lastly, a novel strategy for the FLA-based scheme that min-imizes the effects of delayed feedback is presented and val-idated.
The rest of the paper is structured as follows. Section 2describes the system model under consideration. Section 3briefly reviews the two adaptive schemes covered in thiswork. In Sect. 4 the analytical framework used to analyzethe system goodput is presented. In Sect. 5 the simulationtool features are described and the numerical results com-paring the performance of open- and closed-loop schemesare presented under different configurations and access tech-niques. Finally, in Sect. 6, the main conclusions of this studyare summarized.
2 System overview
2.1 Physical layer description
Our study focuses on the IEEE 802.11n standard [1], whosePHY layer is based on MIMO-OFDM. The MIMO com-ponent enables the use of different transmission techniques(e.g., space-time block coding (STBC), space division mul-tiplexing (SDM), cyclic delay diversity (CDD) and/or com-binations of them) in order to increase the system capac-ity and/or reliability [18]. At the transmitter side, informa-tion bits are first encoded with a 1
2 rate convolutional en-coder with generator polynomials g = [133,171] (in oc-tal form) and then punctured to one of the possible codingrates Cm ∈ {1/2,2/3,3/4,5/6}. Depending on the selectedMIMO configuration, the resulting bits are demultiplexedinto Ns spatial streams. For each stream, the coded bits areinterleaved and then assigned to symbols from one of the al-lowed signal constellations (BPSK, QPSK, 16-QAM or 64-QAM). According to the chosen MIMO mode, the symbolsare then either STBC encoded or antenna mapped on theavailable NT transmit antennas. The resulting symbols arefinally supplied to a conventional OFDM modulator consist-ing of an IFFT (inverse fast Fourier transform) and the ad-dition of a guard interval. For simplicity of exhibition andwithout loss of generality, this paper focuses on a 2 × 2MIMO system (NT = NR = 2), implying that MCSs withNs = 1 and Ns = 2 spatial streams employ STBC [19] andSDM [20], respectively.
At the receiver side, Alamouti decoding or MinimumMean Square Error (MMSE) detection is applied depend-ing on whether STBC or SDM has been employed. Ineither case, the detector extracts soft information in theform of log-likelihood ratios (LLRs) that, after suitable de-interleaving/de-parsing, can be exploited by a soft Viterbidecoder [14].
Modeling fast link adaptation-based 802.11n distributed coordination function 217
Fig. 1 MAC and PPDU frame formats
2.2 MAC layer description
The IEEE 802.11 standard specifies three different MACmechanisms for WLANs, namely the DCF, the point coor-dination function (PCF) and the hybrid coordination func-tion (HCF). The DCF is the mandatory MAC mechanismfor the IEEE 802.11 standard [1] whose frame formats aredetailed in Fig. 1. It is a random access scheme based onthe carrier sense multiple access with collision avoidance(CSMA/CA) protocol that incorporates a binary exponen-tial backoff (BEB) algorithm to manage the retransmissionof collided and erroneous packets.
DCF defines two access techniques, the Basic Access andthe RTS/CTS whose frame exchange is presented in Fig. 2.The Basic Access technique is the most extensively used[3]. In Basic Access, a station (STA) transmits one datapacket at its BEB scheduled slot and waits for its packet-acknowledgement (ACK-control frame) from the receiver(see Fig. 2(a)). If no reply arrives during a predefined timeinterval, the STA interprets the transmission as erroneousand the packet is either retransmitted in the next BEB sched-uled slot or discarded if the number of packet retransmis-sions exceeds the maximum number of allowed retransmis-sions, which will be denoted by R.
In contrast, the implementation of RTS/CTS is optional,yet it is an advisable feature whenever the system operateswith packets whose length exceeds a predefined threshold(Sthres ) or the system has to cope with the hidden terminalproblem [21]. Unlike Basic Access, prior to the data trans-mission, RTS/CTS exchanges two control frames, requestto send (RTS) and clear to send (CTS), between source anddestination (see Fig. 2(b)). This frame exchange allows thereservation of the channel for the current data transmissionand drastically reduces the eventual collision duration to thatof two collided RTS frames. However, RTS/CTS produces aconsiderable increase of the system overhead, lowering inthis way the system performance when short packet lengthsor high transmission rates are used. Nevertheless, thanks toits reduced collision duration, it outperforms Basic Accessin dense user scenarios where the collision probability ishigh.
Apart from DCF, the standard also defines PCF as an op-tional MAC mechanism, which is only usable on infrastruc-ture network configurations. It is a centralized MAC proto-col where a point coordinator (PC), usually the access point,indicates which STA has currently the right to access the
Fig. 2 DCF access mechanisms
medium. This function achieves the collision free operation,using a carrier sense mechanism aided by an access prioritytechnique. However, it is not commonly supported by cur-rent devices and therefore, its performance evaluation willnot be addressed in this paper.
Furthermore, the HCF mechanism, incorporated into thestandard in the IEEE 802.11e amendment, improves DCFand PCF functions with enhanced QoS support, by defin-ing the HCF controlled channel access (HCCA) and theenhanced distributed channel access (EDCA) for PCF andDCF, respectively. Both HCCA and EDCA mechanismsclassify the traffic under different categories, thus providingdifferent access priorities to the medium. The EDCA proto-col is regarded as an enhanced version of DCF and, althoughout of the scope of this paper, it constitutes an interestingthread for future research.
2.3 Frame format novelties of IEEE 802.11n
The IEEE 802.11n amendment has introduced some frameformat changes with respect to previous versions that exploitthe availability of a feedback channel by means of new trans-mission techniques. Often, however, these enhancements areat the cost of some overhead increment and sacrificing back-ward device compatibility. At the MAC layer, the MAC pro-tocol data unit (MPDU) frame format (Fig. 1(a)) includes,among others, the high throughput (HT) control field that al-lows the MCS feedback between transmission entities. Ad-ditionally, two extra PLCP protocol data unit (PPDU) frame
218 G. Martorell et al.
formats are defined at the physical layer with longer pream-ble fields improving the accuracy of the channel estimationand enabling additional transmission techniques (e.g., beam-forming, SDM or CDD). In this work, the PPDU 802.11nframe format presented in Fig. 1(b) has been employed.
2.4 Timing of DCF events
2.4.1 Basic access technique
According to the Basic Access scheme of DCF, the elapsedtime for a successful transmission of an L-bit MPDU usingMCS m is
T bass (m,L) = TT r(m,L) + tSIFS
+ TACK+HT C(m) + tDIFS, (1)
where tSIFS (short interframe space), tDIFS (distributed in-terframe space) and σ is the idle slot duration are 802.11ntime constants defined in [1]. The time elapsed in the MPDUtransmission, TT r(m,L), is defined as
TT r(m,L) = tP reamble + NSym(m,L)tSym, (2)
with tP reamble representing the PLCP preamble duration,tSym denoting the OFDM symbol period and
NSym(m,L) = mST BC
⌈L + 22
mST BCNDBPS(m)
⌉, (3)
being the number of OFDM symbols involved in the trans-mission of a complete MPDU, where NDBPS(m) is thenumber of bits forming each OFDM symbol as defined byMCS m, �z� denotes the smallest integer greater than orequal to z, and mST BC = 2 if STBC is used and mST BC = 1otherwise. Similarly, the time required for the transmissionof an ACK+HTC frame1 using PHY mode m is given by
TACK+HT C(m) = tP reamble + NSym(m,160)tSym. (4)
A collision occurs whenever two or more STAs trans-mit on the same slot, finishing tEIFS (extended interframespace) after the end of the longest transmission of the col-lided STAs. That is, its duration depends on the MCS andMPDU length corresponding to the longest transmission,denoted by m∗ and L∗, respectively. Therefore, the collisionduration can be mathematically expressed as
T basc
(m∗,L∗) = TT r
(m∗,L∗) + tEIFS + σ, (5)
where
tEIFS = tSIFS + TACK(m = 0) + tDIFS. (6)
1Wrapper control frame that encapsulates the ACK and the HT controlfield required to feedback the MCS selection.
Finally, the MPDU error transmission duration, definedas Te(m,L), is the time elapsed in a transmission that ex-periences errors without collisions, and it can be expressedas
T base (m,L) = TT r(m,L) + tEIFS + σ. (7)
2.4.2 RTS/CTS access technique
Unlike Basic Access, the RTS/CTS mechanism incorporatesa frame exchange prior to the data transmission. One of theconsequences of this handshake procedure is that when acollision takes place, its duration is minimized. This is be-cause collisions can only occur during the RTS/CTS ex-change, which only involves frames of minimal length. Inthis paper, a short retry limit (R = 4) has been applied toboth the RTS frames and data packets.
The RTS/CTS time elapsed for a successful transmissionof an L-bit MPDU using MCS m is
T rtss (m,L) = TRT S+HT C(m) + TCT S+HT C(m)
+ TT r(m,L) + TACK+HT C(m)
+ 3tSIFS + tDIFS, (8)
where
TRT S+HT C(m) = tP reamble + NSym(m,160)tSym (9)
and
TCT S+HT C(m) = tP reamble + NSym(m,208)tSym (10)
are control wrapper frames encapsulating RTS and CTSframes, respectively.
Similarly, the RTS/CTS elapsed time in a transmissionerror of an L-bit MPDU using MCS m is
T rtse (m,L) = TRT S+HT C(m) + TCT S+HT C(m)
+ TT r(m,L) + 2tSIFS + tEIFS + σ. (11)
Finally, the collision duration can be defined as
T rtsc
(m′) = tRT S+HT C
(m′) + tEIFS + σ, (12)
where m′ is the lowest MCS value that causes the longestRTS transmission duration from the collided users.
In this model, due to its negligible probability of oc-currence, we have not considered the possibility of an er-ror in the ACK, RTS and/or CTS transmissions. The ACKtransmission takes place under the same system conditionsthan the packet being acknowledged, i.e., using the sameMCS and suffering similar channel conditions. However, itspacket size is considerably smaller than that of the informa-tion packets and therefore, its error probability can be safelyconsidered insignificant. Similarly, packet sizes for RTS andCTS transmissions are also small and the use of the most re-liable MCS warrants a negligible error probability.
Modeling fast link adaptation-based 802.11n distributed coordination function 219
3 Adaptive modulation and coding strategies
3.1 ARF
This algorithm adapts the transmission rate according to thenumber of consecutive transmission failures and successes,both reported by the ACK mechanism. The transmissionrate is decreased after two consecutive transmission failuresand increased after either ten consecutive successful packettransmissions or a timeout. In order to improve the systemadaptation during long intervals of inactivity, this timeoutcounter is reset after a transmission rate change or after atransmission failure [2]. Acceptable timeout values lie inthe range of 50–200 ms [22]. Note that, following a rate in-crease, the next data transmission is deemed as a probingtransmission for the new mode. If an ACK is not receivedfor this probing packet the system falls back to the previousdata rate.
In order to implement ARF in IEEE 802.11n it is neces-sary to determine the available rates in the MCS set, denotedby M. In contrast to previous IEEE 802.11 standards, in802.11n different MCSs ∈ M can provide the same trans-mission rate, but only one of them can be used by the ARFalgorithm. For this reason, the MCSs in M are reorderedaccording to their transmission rate [14], and for those ratesthat can be attained using either SDM or STBC, only theSTBC MCS is kept as it can be shown to be more robustagainst channel variations [19]. The reordered and prunedMCS set will be denoted by M.
3.2 FLA
Fast link adaptation is a closed-loop technique that relies onthe availability of a feedback channel from the receiver tothe transmitter. The main idea behind FLA is that the re-ceiver, thanks to an accurate knowledge of the channel re-sponse, can compute a reliable prediction of the error ratefor all available MCSs and choose the one maximizing theinstantaneous throughput while satisfying QoS constraintsin the form of outage packet error rate probability. The se-lected MCS can then be communicated to the transmitter viathe feedback channel. In this work we assume the use of themethodology presented in [14], where link performance pre-diction for each MCS is based on the exponential effectiveSNR mapping (EESM). Using this approach, the EESM fora given MCS can be easily associated to packet error rate(PER) using look-up tables that have been previously com-puted during an off-line calibration phase.
3.2.1 Basic Access
In crowded scenarios using Basic Access, a high delay be-tween the MCS selection at the receiver and its use at the
Fig. 3 Block diagram of FLA operation with outdated MCS informa-tion
transmitter can be expected, very often exceeding the chan-nel coherence time and significantly affecting the FLA oper-ation. In DCF, all STAs have an equal long term probabilityof accessing the medium. Therefore, successive transmis-sions from a given STA are intertwined with transmissionsfrom the other contending STAs, consequently, the time be-tween successive transmissions increases affecting the MCSfeedback delay. This delay becomes critical for FLA when itexceeds the channel coherence time, indicating that the pro-vided MCS has been determined for a channel response thatis almost uncorrelated to the current channel response. Thismismatch between current and prior channel states can in-crease the error probability due to a mistakenly selected orexpired MCS, causing several consecutive errors in the nextretransmissions prior to packet discard.
In order to counteract the effects of using stale feedbackMCSs, we propose that the STA decreases the transmissionmode when the MCS feedback delay exceeds a fixed timeoutas shown in Fig. 3. The STA will decrease again the MCSin all the subsequent packet retransmissions (if any) untilthe packet is successfully transmitted. The timeout is config-ured to a value close to the channel coherence time, assuringin this way that the current channel response is similar tothe one that the receiver has used to determine the feedbackMCS. As it will be shown in the numerical results section,this strategy reduces the error probability without consider-ably harming neither the goodput nor the fulfillment of QoSconstraints.
3.2.2 RTS/CTS
Unlike Basic Access, in the RTS/CTS scheme the providedMCS is calculated using the channel response affecting the
220 G. Martorell et al.
RTS frame and is returned to the sender in the CTS frameresponse, just before its use for the data transmission. There-fore, the MCS feedback delay is almost negligible comparedto the channel coherence time, and independent of the timeelapsed since the last successful transmission.
4 Goodput analysis
Following the model presented in [10], and recently refinedin [15] and [16], the goodput analysis presented in this paperis suitable for Basic and RTS/CTS access techniques, focus-ing on the saturation region, defined as the operation pointwhere each STA has always new packets to transmit. Thesystem saturation goodput S can be defined as
S = E{payload information in a slot}E{duration of a slot} , (13)
where E{·} denotes statistical expectation. The duration ofa slot refers to the time interval between two consecutivebackoff counter decrements [15].
In any given slot, one out of four events can occur: a suc-cessful packet transmission (s), an error packet transmission(e), a collision (c) or an idle slot (i). From the point of viewof the BEB algorithm, error transmissions and collisions areindistinguishable. The conditional probability of the unionof these events can be computed as
p = 1 − (1 − ζu)(1 − τ)n−1, (14)
where n is the number of active STAs in the scenario, ζu
is the user error transmission probability obtained by sim-ulation for the considered AMC algorithm, averaged acrossusers, and τ is the stationary probability that a particularSTA transmits in a given slot. This transmission probabilitycan be obtained as
τ = 2(1 − pR+1)
(1 − p)Θ, (15)
with
Θ =R∑
j=1
(Wj + 1)
[ζup
j−1 + (W − 1)pj
W + ζu − 1
]+ W + 1
−(
1 − 1 − ζu
W + ζu − 1
)(1 − pR+1), (16)
where Wj = 2min(j,mmax)(CWmin + 1) − 1, CWmin is theminimum contention window, mmax = log2(
CWmax+1CWmin+1 ) de-
noting the maximum backoff stage, CWmax is the maximumcontention window size and W = CWmin + 1. Notice thatp and τ can be obtained by solving the nonlinear systemformed by (14) and (15).
Using τ , the probability that only one STA transmits ona given slot achieving a successful transmission is
Pu = nτ(1 − τ)n−1(1 − ζu). (17)
Similarly, the probability that only one STA transmits on agiven slot achieving an erroneous transmission is
Pe = nτ(1 − τ)n−1ζu. (18)
Furthermore, the probability that a given slot is idle isgiven by
Pi = (1 − τ)n. (19)
Among all possible events, only the successful packettransmission increases the payload information while anyother event leads to a goodput degradation. Consequently,modifying the goodput expression in [16, Eq. (5)], and tak-ing into account the possible use of multiple transmissionmodes at each event, the system goodput can be expressedas
S = PsLp
Piσ + PsTs + PeTe + (1 − Pi − Ps − Pe)Tc
, (20)
where Lp = E{Lp} WW+ζu−1 , with Lp = L−Lh representing
the packet payload length and Lh denoting the MAC sub-layer overhead, and the time values Ts , Tc and Te representthe average elapsed time for successful, colliding and er-ror transmissions, respectively. It should be pointed out thatthese average time values are determined by simulation anddepend on the mean MPDU length (L), the number of users(n) and the probability of use of each MCS according to theAMC employed and the scenario configuration. Notice thatE{Lp} is multiplied by [ W
W+ζu−1 ] in order to account for the
additional information transmitted during anomalous slots2
[15] and Ts = W+ζu
W+ζu−1E{Ts(m,L)}+σ includes the anoma-lous slot duration.
The necessity of obtaining these variables (ζu, Ts , Tc, Te
and Lp) from numerical simulation is what renders our pro-posed model as semi-analytical.
5 Numerical results
In order to validate our semi-analytical model and comparethe performance of FLA and ARF under different systemconfigurations, an IEEE 802.11n system-level Matlab simu-lator has been implemented using the link-level parametersderived in [14] and in accordance with the procedures de-scribed in [1]. Simulator accuracy has been contrasted by
2Slot with a lower probability to be accessed than the average (see [15]for more details).
Modeling fast link adaptation-based 802.11n distributed coordination function 221
Fig. 4 Semi-analytic and simulated system performance using R = 4 and R = 7 retransmissions
reproducing results from [10, 15, 16] for single-mode trans-mission. Note that this model is considerably more realisticthan the one proposed by Tinnirello et al. in [15] and Chen etal. in [16], since it allows the treatment of AMC, statisticalpacket length distribution and non-ideal closed-loop FLAstrategies, at the expense of relying on some semi-analyticparameters.
In this paper we concentrate on the performance evalua-tion of the uplink scenario where, nevertheless, MAC con-trol frame transmissions from access point (AP) to STA arealso accounted for. Different scenarios have been generatedby uniformly distributing n static users in a circular areaof radius Rmax centered on the AP and then determiningthe individual channel response from each user to the AP.To this end, the MIMO channel generation tool presentedin [23], parameterized with each user’s distance to the AP,has been employed. The maximum radius Rmax has beenset to 30 m, a value that ensures the avoidance of the hid-den terminal problem [21] and precludes the utilization ofthe no transmission mode (available in FLA). For all STAs,transmit power has been set to 20 dBm and receiver noisepower to −80 dBm. The physical layer uses only the first16 MCS modes of IEEE 802.11n (MCS0-MCS15), achiev-ing data rates of up to 130 Mbps [1]. It should be stressedthat all the users inside the scenario use the same DCF ac-cess technique, Basic Access or RTS/CTS. The ARF time-out has been set to 60 ms and the FLA outage constraintfor a PER objective (not including collisions) of 10−1 hasbeen configured to 10 %. The corresponding CSI feedbackoverhead has also been taken into account in FLA. In orderto obtain an accurate estimate of the average system perfor-mance, Nsim = 100 simulation runs of duration tsim = 22
seconds for each number of users (n) have been executedusing Matlab.
5.1 Basic Access
Figures 4(a) and 4(b) show the goodput performance andthe conditional probabilities p and τ , respectively, as a func-tion of the number of STAs. Results have been obtained fora fixed packet length of Lp = 1500 bytes and a maximumnumber of allowed retransmissions equal to either R = 4 orR = 7. A very accurate match between the semi-analyticaland simulated system performance metrics for FLA- andARF-based schemes can be appreciated. Figure 4(a) also re-ports the goodput performance obtained using the proposedsemi-analytical model compared to the previous proposal3
presented in [17], where the anomalous slot performanceand the packet retry limit were not considered. Althoughthe previous model provides valuable approximations to thesimulation performance, the new semi-analytical frameworkresults in improved modeling accuracy, especially when thesystem uses R = 4. Figure 4(a) also illustrates that, regard-less of the retry limit, FLA-based schemes outperform ARF-based strategies in terms of goodput performance. Note thatFLA and ARF-based strategies with R = 7 outperform theircounterparts with R = 4 in terms of goodput because a lowerretry limit leads to a lower average backoff contention win-dow, thus increasing the transmission probability (τ ) and,consequently, the collision and error probability (p), seeFig. 4(b).
3Configured to mmax = 4 or mmax = 6 in order to be compared to thenew model using mmax = 6 with R = 4 or R = 7, respectively.
222 G. Martorell et al.
Fig. 5 Goodput, Jain’s fairness index and PER of ARF and FLA strategies using R = 7
Figures 5(a), 5(b) and 5(c), show the goodput, the fair-ness index4 and the PER system performance, respectively,of FLA under the assumptions of ideal and non-ideal chan-nel state information (CSI) for different timeout values (tout )when using R = 7. Under the constraint of maximum sys-tem fairness, the ideal FLA can be considered as the bench-mark system from a goodput point of view. Nevertheless,ideal FLA is not implementable due to the 802.11n MCSfeedback mechanism, which invariably introduces some de-lay in its transmission. In order to improve the FLA perfor-mance for those STAs that experience large MCS feedbackdelays, the FLA algorithm proposed in this paper lowers theMCS rate after the expiration of a finite timeout. Althoughthe timeout should be set according to the channel coherencetime, experimental results show that FLA with tout = 60 msperforms similarly to ideal FLA in terms of goodput, whilepreserving a high fairness index and satisfying PER-basedQoS constraints (see Fig. 5(b) and Fig. 5(c), respectively).
In Fig. 5(a), a priori, ideal FLA could be expected to pro-vide the maximum goodput, however it is outperformed byFLA with tout > 60 ms when n > 5. Nevertheless, as it canbe observed in Fig. 5(b), this goodput improvement is at theexpense of a loss in the Jain’s fairness index measured interms of the per-STA transmission opportunity. This loss infairness reflects that some stations are able to transmit morefrequently than the others. These STAs are the ones whose
4The Jain’s fairness measure used in this paper is calculated as I =(∑n
i βi )2
n∑n
i β2i
where βi denotes the average number of transmissions for
STA i. Note that I = 1n
implies an unfair system and I = 1 reflectsa completely fair system.
channel conditions are so good (high SNR) that the MCSfeedback delay has a negligible effect on their transmissionerror probability (see Fig. 6 where the probability of usingeach transmission rate in a successful transmission is de-picted for different FLA configurations). Consequently, theircontention window is mostly doubled due to collisions andrarely due to erroneous transmissions. In contrast, the otherstations, which are affected by the MCS feedback delay,experience a considerably higher probability of error (seeFig. 5(c)). Accordingly, their DCF mechanism doubles theircontention window after each erroneous transmission result-ing in a lower transmission probability when compared withthat of the STAs with good channel status. The combina-tion of these two types of STAs leads to a system goodputhigher than that of the ideal FLA because stations experienc-ing good channels, which use the highest transmission rates,have more chances of accessing the medium at the expenseof the STAs with poor channel conditions.
In Fig. 7, the performance of FLA and ARF is shownfor a packet length (Lp) modeled as a doubly truncated ex-ponential distribution between 40 and 10.000 bytes. Whenusing FLA, it is assumed that the receiver knows the packetlength to be used for the next transmission according to thedetermination of the most appropriate MCS to use. This as-sumption is quite realistic since there exists a high Lp corre-lation between consecutive packets sent from the same STAin typical WLAN environments.
Figure 7(a) presents the goodput performance of ARF,ideal FLA and FLA (FLA with tout = 60 ms), for differentaverage packet sizes (Lp). Due to the large overhead intro-duced by the DCF mechanism, the adoption of long Lp val-ues improves the DCF protocol efficiency and consequently,
Modeling fast link adaptation-based 802.11n distributed coordination function 223
Fig. 6 Probability of use of the different transmission rates given that the packet has been successfully received. Results are shown for differentFLA settings using a system configuration with R = 7
the system goodput increases, especially for the FLA cases.Note that FLA is still outperforming ARF for any Lp andnumber of users, most notably for those cases where morethan two users are contending for the medium. As previouslyobserved, the goodput performance of FLA with tout = 60ms is similar to that of ideal FLA for the whole range ofLp values and number of active STAs under consideration(see Fig. 7(a)). Furthermore, it keeps Jain’s fairness indexhigh (see Fig. 7(b)) and fulfills the PER QoS constraint forall the considered configurations (see Fig. 7(c)). Notice thatthe system PER performance of FLA increases for long Lp
as a consequence of the obvious increment of the averageMCS feedback delay. For completeness, Figs. 7(b) and 7(c),also present the Jain’s fairness index and PER performance,respectively, for ARF and ideal FLA.
5.2 RTS/CTS
In Fig. 8, FLA and ARF system performance is presentedfor the Basic and RTS/CTS access techniques. ARF onRTS/CTS access is equivalent to the CARA-RTS algorithm
defined in [3], with Pth = 0 (probing activation thresh-old) and Nth = 1 (consecutive failure threshold). AlthoughRTS/CTS is used, FLA is still outperforming ARF in termsof goodput, irrespective of the number of users in the sys-tem. However, the goodput improvement is significantlylower than that obtained in the Basic Access case (seeFig. 8(a)), as now ARF is capable of distinguishing thesource of errors (collisions or channel errors). Note that,due to the reduced MCS feedback delay when employ-ing RTS/CTS, FLA optimally selects the MCS and re-sults in PER values similar to those obtained when us-ing ideal FLA (see Figs. 8(c) and 5(c)). Remarkably, FLAwith Basic Access is outperformed by FLA with RTS/CTSfor n > 15, due to the better performance of RTS/CTS indense environments with high collision probabilities. Fi-nally, it should be mentioned that, as in Basic Access, bothAMCs with RTS/CTS maintain a high degree of fairnessirrespective of the number of users in the system (n) (seeFig. 8(b)).
Figure 9 shows analytical and simulation goodput perfor-mance for different n and Lp , demonstrating the accuracy of
224 G. Martorell et al.
Fig. 7 Goodput, Jain’s fairness index and PER system performance as a function of Lp and n
Fig. 8 FLA and ARF performance using RTS/CTS and Basic Access with fixed Lp = 1500 Bytes
the semi-analytical model and the clear superiority of FLAwith respect to ARF irrespective of the access technique.Moreover, FLA on RTS/CTS, thanks to the reduction in thecollision duration, outperforms FLA on Basic Access forlong packet lengths and n > 2 (see Fig. 9(a)). Lastly, noticethat ARF on RTS/CTS shows an improvement over ARF onBasic Access in terms of goodput for any Lp when n > 2,demonstrating that the overhead introduced by the RTS/CTS
can be tolerated in order to provide reliable information toARF.
In case of a mixture of basic- and RTS/CTS-accessschemes, and depending on the scenario configuration(number of users using each access scheme, packet size oruser signal to noise ratio (SNR)), the throughput results willlie somewhere between the performance of Basic Accessand RTS/CTS. In this case, and thanks to the DCF scheme,
Modeling fast link adaptation-based 802.11n distributed coordination function 225
Fig. 9 System Goodput, as a function of n and Lp , for FLA and ARF adaptation strategies when used with Basic and RTS/CTS access schemes
the system will preserve the medium access fairness be-tween STAs. Remarkably, in this case collisions involvingusers of Basic Access and RTS/CTS access scheme wouldappear, and their duration would correspond to the longesttransmission that would normally correspond to the BasicAccess STAs.
6 Conclusions
This paper has presented a semi-analytical framework forthe performance modeling of MIMO-OFDM WLANs whenusing the Basic Access or RTS/CTS at the MAC layer. Un-like previous work, the proposed model is able to incor-porate the effects of channel errors, the possibility of us-ing open- or closed-loop transmission mode adaptation, theeffect of the retry limit at the MAC layer and the use ofoutdated MCS feedback information. A complete study ofFLA over 802.11n PHY/MAC in terms of goodput, fairnessand system PER performance for a wide range of numberof users and packet sizes has been presented and contrastedwith those obtained using ARF. Noteworthy, the influence
of feedback delay on the performance of FLA has beenassessed. In Basic Access scheme, the degradation causedby an outdated MCS information can be largely compen-sated with the use of a time-out strategy that weighs downthe influence of the MCS feedback delay. Numerical resultsclearly show that as the number of users in the system grows,the FLA-based adaptation proves to be much more robustto collisions than ARF even when employing outdated MCSfeedback information. This effect is clearly demonstrated bythe fact that whereas ARF-based schemes suffer a dramaticreduction in goodput for more than two users, the FLA-based strategy exhibits a very graceful degradation thanks toa more accurate rate selection in the presence of collisions.
When using the RTS/CTS access scheme, FLA still beatsARF irrespective of the number of users and packet sizes,although the difference is not as significant as in Basic Ac-cess. Nevertheless, it should be remarked that the RTS/CTSframe exchange allows a delay-free selection of the MCSand results in a system PER similar to the one obtained in theideal FLA case. Overall, it can be concluded that FLA yieldsa goodput that significantly outperforms ARF for most sys-tem loads and access schemes, while keeping a large degree
226 G. Martorell et al.
of fairness and satisfying prescribed PER-based QoS con-straints.
As future work, the performance of FLA under non-saturated traffic conditions will be assessed. Additionally,the performance of FLA when jointly implemented withcarrier sense multiple access with enhanced collision avoid-ance (CSMA/E2CA) and/or under multiuser multiple-inputmultiple-output (MU-MIMO) configurations will also bestudied. Finally, different techniques able to counteract feed-back delays surpassing the coherence time must be devised.
References
1. IEEE (2009) IEEE Std 802.11n-2009, Part 11: Wireless LANmedium access control (MAC) and physical layer (PHY) speci-fications amendment 5: Enhancements for higher throughput.
2. Kamerman, A., & Monteban, L. (1997). WaveLAN®-II: a high-performance wireless LAN for the unlicensed band. Bell LabsTechnical Journal, 2(3), 118–133.
3. Kim, S., Verma, L., Choi, S., & Qiao, D. (2010). Collision-awarerate adaptation in multi-rate WLANs: design and implementation.Computer Networks, 54(17), 3011–3030.
4. Joshi, T., Ahuja, D., Singh, D., & Agrawal, D. (2008). SARA:stochastic automata rate adaptation for IEEE 802.11 networks.IEEE Transactions on Parallel and Distributed Systems, 19(11),1579–1590.
5. He, J., Kaleshi, D., Munro, A., & McGeehan, J. (2006). Model-ing link adaptation algorithm for IEEE 802.11 wireless LAN net-works. In IEEE ISWCS, Valencia, Spain, Sept. 2006.
6. Jung, H., Kwon, T., Choi, Y., & Seok, Y. (2007) A scalable rateadaptation mechanism for IEEE 802.11e wireless LANs. In IEEEFGCN, Jeju-Island, Korea, Dec. 2007.
7. Zhang, J., Tan, K., Zhao, J., Wu, H., & Zhang, Y. (2008). A prac-tical SNR-guided rate adaptation. In IEEE INFOCOM, Phoenix,AZ, April 2008.
8. Holland, G., Vaidya, N., & Bahl, P. (2001). A rate-adaptive MACprotocol for multi-hop wireless networks. In ACM MobiCom (pp.236–251).
9. Choi, J., Na, J., sup Lim, Y., Park, K., & kwon Kim, C. (2008).Collision-aware design of rate adaptation for multi-rate 802.11WLANs. IEEE Journal on Selected Areas in Communications,26(8), 1366–1375.
10. Bianchi, G. (2000). Performance analysis of the IEEE 802.11 dis-tributed coordination function. IEEE Journal on Selected Areas inCommunications, 18(3), 535–547.
11. Park, C., Han, D., & Ahn, S. (2006). Performance analysis ofMAC layer protocols in the IEEE 802.11 wireless LAN. Telecom-munications Systems, 33, 233–253.
12. Szczypiorski, K., & Lubacz, J. (2008). Saturation throughput anal-ysis of IEEE 802.11g (ERP-OFDM) networks. Telecommunica-tions Systems, 38, 45–52.
13. Martorell, G., Riera-Palou, F., & Femenias, G. (2009). Cross-layerlink adaptation for IEEE 802.11n. In IEEE IWCLD, Palma, Spain,June 2009.
14. Martorell, G., Riera-Palou, F., & Femenias, G. (2011). Cross-layerfast link adaptation for MIMO-OFDM based WLANs. WirelessPersonal Communications, 56(3), 599–609.
15. Tinnirello, I., Bianchi, G., & Xiao, Y. (2010). Refinements onIEEE 802.11 distributed coordination function modeling ap-proaches. IEEE Transactions on Vehicular Technology, 59(3),1055–1067.
16. Chen, H. (2011). Revisit of the Markov model of IEEE 802.11DCF for an error-prone channel. IEEE Communications Letters,15(12), 1278–1280.
17. Martorell, G., Riera-Palou, F., & Femenias, G. (2011). DCFperformance analysis of open- and closed-loop adaptive IEEE802.11n networks. In IEEE ICC, Kyoto, Japan, June 2011.
18. Goldsmith, A. (2005). Wireless communications. Cambridge:Cambridge University Press.
19. Choi, Y.-S., & Alamouti, S. (2008). A pragmatic PHY abstractiontechnique for link adaptation and MIMO switching. IEEE Journalon Selected Areas in Communications, 26(6), 960–971.
20. Foschini, G. (1996). Layered space-time architecture for wirelesscommunication in a fading environment when using multi-elementantennas. Bell Labs Technical Journal, 1(2), 41–59.
21. Ghaboosi, K., Latva-aho, M., & Pomalaza-Ráez, C. (2008). Anovel MAC protocol and layer two transmission scheduling al-gorithm for WLANs. Telecommunications Systems, 37, 3–18.
22. Holland, G., Vaidya, N., & Bahl, P. (2001). A rate-adaptive MACprotocol for multi-hop wireless networks. In ACM MobiCom,Rome, Italy.
23. Kermoal, J., Schumacher, L., Pedersen, K., Mogensen, P., & Fred-eriksen, F. (2002). A stochastic MIMO radio channel model withexperimental validation. IEEE Journal on Selected Areas in Com-munications, 20(6), 1211–1226.
Gabriel Martorell was born in1984 in Porreres, Spain. He re-ceived the B.S. degree in Telecom-munication Engineering from theUniversity of the Balearic Islands(UIB), Spain, in 2006 and the M.S.degree in Telecommunication En-gineering from Technical Univer-sity of Catalonia (UPC), Spain, in2008. He is currently working to-wards the Ph.D. degree in the mo-bile communications group at UIB,with funding from the governmentof the Balearic Islands. His mainresearch interests are mobile and
wireless communications with an emphasis on adaptation techniquessuitable for future wireless communications systems. He is a researchmember of AM3DIO (Spanish government project).
Felip Riera-Palou was born in1973 in Palma, Mallorca (Spain).He received the BS/MS degree inComputer Engineering from theUniversity of the Balearic Islands(UIB), (Mallorca, Spain) in 1997,the MSc and PhD degrees in Com-munication Engineering from theUniversity of Bradford (UK) in1998 and 2002, respectively, andthe MSc degree in Statistics fromthe University of Sheffield (UK) in2006. From May 2002 to March2005, he was with Philips Re-search Laboratories (Eindhoven,
The Netherlands) first as a Marie Curie postdoctoral fellow (EuropeanUnion) and later as a member of technical staff. While at Philips heworked on research programs related to wideband speech/audio com-pression and speech enhancement for mobile telephony. From April2005 to December 2009 he was a research associate (Ramon y Cajalprogram, Spanish Ministry of Science) in the Mobile Communications
Modeling fast link adaptation-based 802.11n distributed coordination function 227
Group of the Dept. of Mathematics and Informatics at UIB. Since Jan-uary 2010 he is an associate research professor (I3 program, SpanishMinistry of Education) at UIB. Dr Riera-Palou’s current research in-terests are in the general areas of signal processing and wireless com-munications. He is a Senior Member of the IEEE.
Guillem Femenias received theTelecommunication Engineer de-gree and the Ph.D. degree in elec-trical engineering from the Techni-cal University of Catalonia (UPC),Barcelona, Spain, in 1987 and 1991,respectively.From 1987 to 1994, he worked asa researcher with UPC, where hebecame an Associate Professor in1990. In 1995, he joined the De-partment of Mathematics and Infor-matics, University of the BalearicIslands (UIB), Mallorca, Spain,where he became Full Professor in
2010. Dr. Femenias is currently leading the Mobile CommunicationsGroup at UIB, where he has been the Project Manager of projectsARAMIS, DREAMS, DARWIN, MARIMBA, and COSMOS, all ofwhich being funded by the Spanish and Balearic Islands Governments.In the past, he was also involved with several European projects (AT-DMA, CODIT, and COST). His current research interests and activitiesspan the fields of digital communications theory and wireless commu-nication systems, with particular emphasis on cross-layer transceiverdesign, resource management, and scheduling strategies applied tofourth-generation wireless networks. On these topics, he has publishedmore than 75 journal and conference papers, as well as some bookchapters.Dr. Femenias is a Senior Member of IEEE. He was the recipient ofthe Best Paper Awards at the 2007 IFIP International Conference onPersonal Wireless Communications and at the 2009 IEEE VehicularTechnology Conference-Spring. He has served for various IEEE con-ferences as a Technical Program Committee Member, as the Publica-tions Chair for the IEEE 69th Vehicular Technology Conference (VTC-Spring 2009), and as one of the Local Arrangement Chairs for the IEEEStatistical Signal Processing Workshop (SSP’16).
