This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
1

Performance Analysis and Enhancement of

Beamforming Training in 802.11ad
Wen Wu, Member, IEEE, Nan Cheng, Member, IEEE, Ning Zhang, Senior Member, IEEE,
Peng Yang, Member, IEEE, Khalid Aldubaikhy, Xuemin (Sherman) Shen, Fellow, IEEE

Abstract—Beamforming (BF) training is crucial to establishing field-trials have paved the road for the success of mmWave
reliable millimeter-wave communication connections between sta- communications. As the first ratified standard on mmWave
tions (STAs) and an access point. In IEEE 802.11ad BF training wireless local area network (WLAN), IEEE 802.11ad can de-
protocol, all STAs contend for limited BF training opportunities,
i.e., associated BF training (A-BFT) slots, which results in severe liver a data rate of 7 Gbit/s [11]. Moreover, the on-going IEEE
collisions and significant BF training latency, especially in dense 802.11ay building on the legacy of 802.11ad, is anticipated to
user scenarios. In this paper, we first develop an analytical model offer a data rate up to 40 Gbit/s [12]–[14].
to evaluate the BF training protocol performance. Our analytical The mmWave communications, unfortunately, suffer from
model accounts for various protocol components, including user severe free-space path loss because of the high-frequency
density, the number of A-BFT slots, and protocol parameters, i.e.,
retry limit and contention window size. We then derive the aver- band [15]. To compensate for the path loss in mmWave com-
age successful BF training probability, the BF training efficiency munications, beamforming (BF) technology, which focuses the
and latency. Since the derived BF training efficiency is an implicit radio frequency power in a narrow direction, is adopted at
function, to reveal the relationship between system parameters both the transmitter and the receiver. Since reliable commu-
and BF training performance, we also derive an approximate nication is only possible when the BF of both the transmitter
expression of BF training efficiency. Theoretical analysis indicates
that the BF training efficiency degrades drastically in dense user and receiver are properly aligned, a BF training procedure
scenarios. To address this issue, we propose an enhancement between transmitter and receiver is required. Without proper
scheme which adaptively adjusts the protocol parameters in BF training, the data rate of mmWave WLAN may drop from
tune with user density, to improve the BF training performance several Gbps to only a few hundred Mbps [16]. Hence, an
in dense user scenarios. Extensive simulations are carried out efficient BF training is imperative for mmWave networks.
to validate the accuracy of the developed analytical model. In
addition, simulation results show that the proposed enhancement Many recent works have investigated BF training schemes,
scheme can improve the BF training efficiency by 35% in dense such as codebook-based beam search [17], compressed sensing
user scenarios. schemes [18], and out-of-band solutions [19]. Although these
works can significantly improve the efficiency of BF training,
Index Terms – Beamforming training protocol, mmWave, IEEE
802.11ad, enhancement scheme, analytical model.
they mainly consider the BF training performance from the
physical layer’s perspective. However, the contention feature
I. I NTRODUCTION of the BF training from the protocol’s perspective is seldom
considered, i.e., multiple stations (STAs) have to contend for
The prevalence of emerging data-hungry services, such as
the limited BF training opportunities, which results in severe
myriad virtual/augmented reality gaming [1], smart city [2],
collisions, wastes the cherished BF training opportunities and
social networking [3], and big data analytics [4], [5], is
incurs extra BF training latency, especially in dense user
fueling the skyrocketing growth of data traffic in the near
scenarios [20]. Hence, the elaborate analysis of the BF training
future [6], [7]. Cisco reported that the next five years would
protocol is an essential process to determine the applicability
witness a seven-fold increase in global mobile data traffic. To
of the BF training protocol in mmWave networks.
tame the data tsunami, millimeter-wave (mmWave) commu-
IEEE 802.11ad specifies a distributed BF training protocol.
nication emerges as the most promising wireless technology
Specifically, the BF training duration is divided into several
that offers a “wire-like” connection by exploiting a swath of
associated beamforming training (A-BFT) slots, and all STAs
spectrum [8]–[10]. Multiple standardization efforts, plethora
will distributely contend for these limited A-BFT slots in
commercial-off-the-shelf (COTS) products, and large-scale
a contention and backoff manner. However, modeling the
W. Wu, and X. Shen are with the Department of Electrical and Com- performance of BF training protocol is a challenging task.
puter Engineering, University of Waterloo, Waterloo, ON N2L 3G1, Canada The reason is three-fold. Firstly, modeling the BF training
(email:{w77wu, kaldubai, sshen}@uwaterloo.ca). Corresponding author:
Peng Yang. protocol should incorporate protocol components to unveil
N. Cheng is with State Key Lab. of ISN, and with the School of the relationship between various protocol parameters, which
Telecommunications Engineering, Xidian University, Xian 710071, China (e- requires understanding and characterizing the protocol. Sec-
mail: [email protected]); N. Zhang is with the Department of Computing
Sciences, Texas A&M University at Corpus Christi, TX, 78412, USA (email: ondly, due to the “deafness” problem caused by high di-
[email protected]); P. Yang is with the School of Electronic Informa- rectionality feature of BF, i.e., an STA may not be aware
tion and Communications, Huazhong University of Science and Technology, of the transmission of other STAs, the BF training protocol
Wuhan, 430074, P.R. China (email: [email protected]); K. Aldubaikhy is
with the Department of Electrical Engineering, Qassim University, Buraydah, is different from traditional carrier sensing based protocols.
52571, Saudi Arabia (e-mail: [email protected]). Hence, existing analytical models for traditional microwave

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
2

WLANs are unsuitable for the BF training protocol. Thirdly, work is reviewed in Section II. An overview of BF training
the explicit mathematical relationship between the system procedure in 802.11ad is presented in Section III. The pro-
parameters (e.g., user density) and BF training performance posed analytical model and the corresponding performance
is crucial to the protocol optimization in dense user scenarios. analysis are given in Section IV. Section V provides approx-
Thus, we argue that a thorough theoretical framework for imate analysis on the BF training efficiency. Then, Section
802.11ad BF training protocol is necessary and significant. VI proposes an enhancement scheme in dense user scenarios.
Moreover, 802.11ad only provides at most eight A-BFT slots, In Section VII, extensive simulations are conducted to validate
and hence the collision probability is high in dense user the proposed analytical model and the enhancement scheme. In
scenarios, which results in high latency, and low BF training Section VIII, concluding remarks are given, and future works
efficiency (defined as the percentage of A-BFT slots that have are discussed.
been successfully utilized). Thus, an enhancement scheme on
802.11ad is required to improve the performance in dense user II. R ELATED W ORK
scenarios. A collection of works are devoted to enhancing the per-
In this paper, we study the performance of the BF training formance mmWave networks [21]–[23]. Yu et al. proposed a
protocol in 802.11ad. Specifically, we target at answering the novel two-stage algorithm, in which transmission clustering
following questions: 1) How good the performance of the BF and path routing are jointly optimized to reduce backhaul
training protocol is; and 2) How to further enhance the BF traffic in the mmWave mesh network [21]. A pioneering work
training protocol performance in dense user scenarios? Firstly, developed a fast packet transmission scheduling scheme based
to evaluate the performance of the BF training protocol, a on an inducing coloring technique, to combat the Rayleigh-
two-dimensional Markov chain analytical model is developed, fading interference in wireless networks [22]. Another inter-
which takes the number of consecutive collisions and the esting work investigated a joint optimization of beam transmit
backoff time as a state. Our analytical result unveils the power and BF gain to adapt to the time-varying traffic in
impacts of the number of A-BFT slots, the number of STAs mmWave satellite communications [23]. The above works
and protocol parameters, i.e., the retry limit and the contention focus on improving network performance from the perspective
window size, on the performance of the BF training proto- of resource management. As a result, they do not consider the
col. Extensive simulation results validate the accuracy of the impact of BF training on mmWave networks. In contrast, our
proposed analytical model. Secondly, based on the analytical work aims at improving mmWave network performance from
model, we derive the expressions of the successful BF training the perspective of BF training.
probability, BF training efficiency and BF training latency. There are extensive research efforts on designing efficient
Thirdly, since the derived expression of BF training efficiency BF training schemes [17]–[19], [24]–[27]. A codebook-based
is an implicit function, to characterize the relationship between search scheme is proposed in [17], in which beamwidth is
system parameters and BF training performance, we derive adjusted in each step until the optimal beam is identified.
an approximate expression of the BF training efficiency in Marzi et al. developed a compressed sensing based BF training
dense user scenarios, which demonstrates that the BF training method with a low complexity by exploiting the sparse char-
efficiency depends on the ratio of the number of STAs to the acteristic of mmWave channels [18]. An out-of-band scheme
number of A-BFT slots. Particularly, the maximum BF training that exploits traditional Wi-Fi signals to reduce BF training
efficiency is 1/e. Finally, since the performance substantially overhead, is developed in [19]. Similarly, another method that
degrades in dense user scenarios due to the limitation of exploits the spatial correlation in low-frequency band radio
A-BFT slots in practical systems, an enhancement scheme signals, is developed to improve BF training performance
which adaptively adjusts protocol parameters according to the in [24]. In another line of research, Hassanieh et al. proposed
user density, is proposed to improve BF training performance. a fast BF training scheme by utilizing the multi-armed beam
Extensive simulations are carried out and demonstrate that the feature of directional antennas [25]. Wu et al. developed a
enhancement scheme can significantly improve the BF training learning-based BF training scheme [26], in which the corre-
efficiency and reduce the BF training latency, as compared lation structure among nearby beams is exploited to speed
to 802.11ad with the default parameter setting. The main up BF training procedure. In addition, Zhou et al. proposed a
contributions in this paper are summarized as follows: fast BF training scheme for the high-mobility unmanned aerial
vehicles mesh networks [27], which can effectively guarantee
• We develop an accurate analytical model to evaluate
the robustness of mmWave communication links. While the
the performance of BF training protocol. Based on the
aforementioned works can enhance the performance of BF
proposed analytical model, we derive the successful BF
training, they lack of the considerations of the contentions of
training probability, BF training efficiency and average
multiple STAs in the BF training protocol. In contrast, our
BF training latency;
work focuses on investigating the BF training performance
• We derive an approximate expression of the BF training
from the protocol’s perspective.
efficiency to elaborate the relationship between system
On the other hand, some recent research works in [28]–
parameters and BF training performance;
[32] are devoted to the BF training protocol. To alleviate
• We propose an enhancement scheme to improve the BF
collisions of BF training in dense user scenarios, a pioneering
training performance in dense user scenarios.
work [28] proposed a secondary backoff scheme in the A-
The remainder of this paper is organized as follows. Related BFT stage, in which each STA selects not only a backoff

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
3

Table I Beacon interval # t-1 Beacon interval # t Beacon interval # t+1

SUMMARY OF MAIN ACRONYMS.
time
Acronyms Description
A-BFT Associated beamforming training BTI A-BFT ATI DTI
AP Access point
ATI Announcement transmission interval
BF Beamforming
A-BFT slot #1 A-BFT slot #2 … A-BFT slot #M
BI Beacon interval
BTI Beacon transmission interval
DCF Distributed coordinate function SSW-
DTI Data transmission interval AP FB
Directional
MAC Medium access control transmit beam

SSW Sector sweep Omni-directional

receive beam
SSW-FB Sector sweep feedback STA SSW SSW … SSW
STA Station

Fig. 1. The 802.11ad beacon interval format and an illustration of BF training

A-BFT slot, but also a secondary backoff time within the A- procedure.
BFT slot such that transmission collisions can be reduced.
Leveraging the channel sparsity in the mmWave channel, Shao STAs. AP coordinates the BF training, link scheduling and
et al. developed a compressed sensing method of performing network synchronization in the network. Both AP and STAs
BF training simultaneously for a group of STAs in [29]. adopt the directional multi-gigabit mode, i.e., each node is
In addition, there are multiple industry efforts on enhancing equipped with an electrically steerable directional antenna
the BF training protocol. Kim et al. in [30] spread out the capable of supporting the directional transmission.
access attempt over time to cope with the high collision To establish reliable mmWave communication links, an
issue in dense user networks. Another draft in [31] allowed STA should perform BF training with AP at the beginning
BF training simultaneously for different STAs over multiple of each beacon interval (BI). As depicted in Fig. 1, the
channels to enhance BF training, which may increase the transmission time is partitioned into multiple BIs. Each BI
protocol signaling overhead. Jo et al. proposed a short sector is further segregated into: a) beacon transmission interval
sweep (SSW) frame structure in [32] which has a shorter (BTI), during which AP performs the BF training with STA;
packet length compared with an SSW frame, to increase the b) A-BFT, during which all STAs contend for BF training
BF training capability in an A-BFT slot. Although these works with AP; c) announcement transmission interval (ATI), which
provide efficient solutions on enhancing BF training, they coordinates the transmission scheduling in data transmission
more or less need to modify the protocol and hence may be interval (DTI); and d) DTI, which facilitates directional data
incompatible with current 802.11ad. In contrast to these works, transmission [11]. Our paper focuses on the BF training in A-
rather than proposing new schemes with distinct features, BFT. A-BFT is further divided into multiple A-BFT slots to
we target on an in-depth understanding of the 802.11ad BF provide separated BF training for different STAs. Specifically,
training protocol. The reason is two-fold. Firstly, building on the BF training procedure between AP and an STA in an A-
the base of successful 2.4/5 GHz WiFi systems, 802.11ad is the BFT slot is illustrated in Fig. 1. An STA transmits multiple
most practical standard in mmWave WLANs, which has been SSW frames via different directional beams, and AP receives
widely used in many COTS mmWave devices. Secondly, the these SSW frames via an omni-directional beam. Based on
on-going 802.11ay is the follow-up of 802.11ad, which aims the received signal strength of these directional beams, AP
to enhance the legacy 802.11ad while guaranteeing backward can identify the best transmit beam of the STA. Subsequently,
compatibility for legacy users [33]. The 802.11ay is expected the AP sends an SSW-feedback (SSW-FB) frame to the
to adopt a similar BF training protocol, with an increased transmitting STA for the acknowledgment of a successful BF
number of A-BFT slots [34]. The number of A-BFT slots is training. Note that an A-BFT slot can only provide a BF
expected to be increased from 8 in 802.11ad to 40 in 802.11ay training opportunity for an STA. If two STAs compete for
to alleviate the BF training collisions. As such, our analytical the same A-BFT slot, a collision occurs. As such, no SSW-
model can be readily extended to study the performance of FB frame would be sent to STAs, and then the transmitting
802.11ay. STAs are aware of the occurrence of the collision.
Note that above mentioned BF training in A-BFT is a part of
III. BF T RAINING IN 802.11 AD the entire BF training procedure in 802.11ad. As we focus on
This section first presents the BF training procedure in the BF training performance from the protocol’s perspective,
802.11ad, and then details the BF training protocol. For better the detailed BF training procedure is beyond the scope of our
understanding, Table I provides a summary of the acronyms paper. For more details, one can refer to a detailed review
used hereinafter. in [12].

A. BF Training Procedure B. 802.11ad BF Training Protocol

We consider a WLAN network compliant with the 802.11ad To coordinate BF training for multiple STAs, 802.11ad
standard, consisting of an access point (AP) and multiple specifies a distributed BF training protocol, namely, in which

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
4

AP SSW-FB Table II
S UMMARY OF NOTATIONS .

Notation Description
STA A SSW SSW … SSW p Conditional collision probability
ps Conditional successful transmission probability
τ The probability that an STA is active
STA B SSW SSW … SSW
C(t) Consecutive collision counter at time t
B(t) Backoff time at time t
collision (r, w) State with r consecutive collisions and w backoff time
M Number of A-BFT slots
STA C
SSW SSW … SSW R Retry limit
A-BFT slot #1 A-BFT slot #2
… time D Average BF training latency
W Contention window size
S BF training efficiency
Fig. 2. Illustration of the BF training protocol. The BF training in A-BFT π Steady state probability vector
slots #1 is successful, while that in A-BFT slot #2 is unsuccessful since this TBI Duration of a beacon interval
A-BFT slot is selected by two STAs simultaneously.
TSSW Duration of a sector sweep frame
Z+ Positive integer set
each STA performs BF training with AP in a contention and F Number of SSW frames in an A-BFT slot
backoff manner. The BF training protocol consists of the
following two steps.
different scenarios. However, compared with the celebrated
• Before BF training, each active STA randomly selects an distributed coordinate function (DCF) protocol in traditional
A-BFT slot from the range [1, M ] for the transmission omni-directional WLANs, the absence of carrier sensing
of BF training packets (referred to as a transmission for mechanism makes the BF training protocol susceptible to
short hereinafter). Here, M denotes the number of A- collision, especially in dense user scenarios. In what follows,
BFT slots. A successful BF training depends on whether we analyze the performance of the BF training protocol.
an A-BFT slot is selected by multiple STAs. If an A-BFT
slot is only selected by one STA, the transmission would
be successful, and an SSW-FB frame would be sent back IV. A NALYTICAL M ODEL AND P ERFORMANCE A NALYSIS
from the AP. As shown in Fig. 2, the BF training of STA In this section, we first develop a two-dimensional Markov
A in A-BFT slot #1 is successful since this A-BFT slot is chain analytical model for 802.11ad BF training protocol.
not selected by any other STAs. Otherwise, if more than Next, the average successful BF training probability and aver-
one STA select the same A-BFT slot, a collision would age BF training latency are derived. In the following analysis,
occur, and no SSW-FB would be sent back from AP, such we assume a fixed number of STAs and perfect physical
as STA B and STA C in Fig. 2. channel conditions (i.e., no transmission errors) with line-
• After BF training, if the number of consecutive colli- of-sight (LOS) communications in the network. The perfect
sions of an STA exceeds retry limit R (referred to as physical channel condition assumption is widely adopted in
dot11RSSRetryLimit in 802.11ad), the STA would select the medium access control (MAC) protocol analysis [35]–
a discrete backoff time uniformly at random from [0, W ). [37]. In perfect channel conditions, the failure of one BF
Here, W denotes the contention window size (referred training attempt can only be caused by packet transmission
to as dot11RSSBackoff in 802.11ad). Specifically, each collision, which simplifies the MAC performance analysis
STA maintains a consecutive collision counter (referred since other factors (e.g., decoding error) do not need to be
to as FailedRSSAttempts in 802.11ad) which indicates considered. The assumption is also reasonable in practical
the number of consecutive collisions that the STA has mmWave systems, since the 802.11ad standard is designed
experienced in A-BFT. Once a collision occurs, the for indoor mmWave communications where the LOS link
consecutive collision counter is incremented by one. Oth- is dominant. As strong signal strength can be observed in
erwise, upon a successful transmission, the consecutive short-distance LOS communications, the transmission error is
collision counter is cleared to zero. negligible. Field measurements indicate an extremely low bit
Note that STAs can only transmit when the backoff time is error rate in LOS communications [38]. Under imperfect phys-
zero. The backoff time is decremented by one after one BI. ical channel conditions, the probability that one BF training
Thus, if the backoff time of an STA is w, the STA has to attempt fails (p) is a joint probability of packet transmission
be frozen from transmission in the subsequent w BIs. Due to collision probability (pc ) due to the MAC protocol and packet
the backoff mechanism, not all STAs are contending for A- error probability (pe ) due to imperfect channel conditions,
BFT slots. We refer to the contending STAs as active STAs, i.e., p = 1 − (1 − pc )(1 − pe ) [39]. With the relationship,
while other STAs whose backoff times are nonzero, are called our proposed analytical model can be easily extended to the
inactive STAs. study of BF training performance under imperfect channel
The advantage of 802.11ad BFT training protocol is salient. conditions. Hence, we have considered a perfect channel
Firstly, BFT training protocol is fully distributed which is condition case in this work to better elaborate the BF training
scalable with the network size. Secondly, the BF training protocol performance. In addition, we assume that each STA
protocol is simple which can be easily implemented under always needs to perform BF training at each BI [37], [40],

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
5

1-p

The STA would be frozen from transmission in A-BFT

0,0 in the subsequent w BIs.
p
• The backoff time is decremented by one after every BI.
1,0 An STA transits from state (R, w) to state(R, w −1), and
the one step transition probability is given by
…
P (R, w − 1|R, w) = 1, ∀w ∈ [1, W − 1]. (4)
p

R-1,0 • The consecutive collision counter will not be incremented

p/W
… if it reaches the retry limit. Upon an unsuccessful trans-
R,0
1
R,1 … 1
R,W-2
1
R,W-1
mission, an STA transits from state (R, 0) to backoff state
… (R, w), and the corresponding transition probability is
p/W
given by
p
P (R, w|R, 0) = , ∀w ∈ [0, W − 1]. (5)
Fig. 3. Two-dimensional Markov chain model for the BF training protocol. W
Hence, the STA remains in backoff states for the subse-
[41]. This assumption is reasonable in dynamic mmWave net- quent w BIs.
works. Since the established communication connections are
Note that states whose backoff times are zero are called
intermittent and short-lived due to the mobility and blockage,
active states, while the other states are referred to as inactive
the BF training procedure would be invoked persistently. For
states. Here, the steady probability of state (r, w) is defined
better illustration, a summary of important notations is given
as
in Table II.
πr,w = lim P (C(t) = r, B(t) = w) (6)
t→∞

A. Markov Model for the BF Training Protocol and the corresponding steady state probabil-
ity of the proposed Markov chain is π =
We consider a discrete time slotted system, where t denotes
{π0,0 , π1,0 , ..., πR−1,0 , πR,0 , πR,1 , ..., πR,W −1 } ∈ R(R+W )×1 .
the beginning of the t-th BI. To evaluate the performance of the
Let P ∈ R(R+W )×(R+W ) represent the state transition matrix
BF training protocol, we examine a tagged STA and represent
whose nonnull elements are given in (1)-(5). Mathematically,
its status by a two-dimensional Markov chain {C(t), B(t)}.
π can be obtained via solving the following balance equations:
Let C(t) ∈ [0, R] denote the number of consecutive collisions
that the tagged STA has experienced. Let B(t) ∈ [0, W − 1]
denote the current backoff time of the tagged STA. For exam- Pπ = π (7a)
ple, state (r, w) indicates that the tagged STA has experienced R W
X X −1
r consecutive collisions and its current backoff time is w. πr,w = 1 (7b)
Let p denote the average collision probability of the tagged r=0 w=0
STA. Note that p is the conditional collision probability since where (7b) accounts for the fact that the summation of all
the collision occurs only when the tagged STA is active. steady state probabilities equals one.
Accordingly, 1 − p denotes the average conditional successful Remark 1: Our proposed analytical model is different from
transmission probability. The state transition diagram is de- the celebrated Bianchi’s model [37] in the following two
picted in Fig. 3, which is governed by the following events ways. Firstly, we have different backoff mechanisms. STAs
and the corresponding one-step transition probabilities. will backoff after every collision in Bianchi’s model, while in
• Upon a collision, the consecutive collision counter is our model STAs only backoff when the consecutive collision
incremented by one when it does not exceed the retry counter exceeds the retry limit. Secondly, the contention win-
limit. The STA transits from state (r, 0) to state (r+1, 0), dow size increases with the number of consecutive collisions
and the corresponding transition probability is given by in Bianchi’s model, while the contention window size is fixed
in our model.
P (r + 1, 0|r, 0) = p, ∀r ∈ [0, R − 2]. (1)
• Upon a successful transmission, the consecutive collision
counter is cleared to zero. The STA transits from state B. Successful BF Training Probability
(r, 0) to state (0, 0) with the following transition proba- Based on our analytical model, we first derive the closed-
bility from expression of the steady state probability and then obtain
P (0, 0|r, 0) = 1 − p, ∀r ∈ [0, R]. (2) the successful BF training probability.
Theorem 1: The steady state probabilities of the proposed
• If the consecutive collision counter reaches retry limit R,
Markov chain can be represented by
the STA randomly selects a backoff time w ∈ [0, W − 1].
The STA transits from state (R − 1, 0) to backoff state pr (1 − p)

 pR (W − 1) /2 + 1 , ∀r ∈ [0, R − 1], w = 0


(R, w), and the corresponding the transition probability
is πr,w = (8)
p  (W − w) pR
P (R, w|R − 1, 0) = , ∀w ∈ [0, W − 1]. (3)

 , ∀r = R, w ∈ [0, W − 1].
W W (pR (W − 1) /2 + 1)

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
6

Proof 1: The detailed proof is given in Appendix A. be represented by

All steady state probabilities in Theorem 1 are represented ∞
X
by p which is the unknown conditional collision probability. D= P (Success|Collisions = i) E [Di ]
In the following, the value of p will be obtained. i=0
(13)
∞
The probability that an STA stays in an active state, is given
X
i
= (1 − p) p E [Di ]
by i=0
R
X 1 where E[Di ] represents the BF training latency of a successful
τ= πr,0 = . (9) transmission after experiencing i consecutive collisions. Note
r=0
pR (W − 1) /2 + 1
that if the number of consecutive collisions exceeds the retry
From the perspective of the STA, a successful transmission limit, the STA would be frozen from transmission for a backoff
only occurs when other active STAs select other A-BFT time. Hence, the calculation of E[Di ] can be divided into
slots for transmission or stay in inactive states. Hence, the the following two cases based on the number of consecutive
conditional successful transmission probability, given an STA collisions:
is active, is
   N −1
1
ps = τ 1 − +1−τ
M
 τ N −1

= 1− (10)
M • When i < R, E [Di ] can be represented by
 N −1
1
= 1− R
. E [Di ] = i · TBI + F · TSSW = TBI (i + α) (14)
M (p (W − 1) /2 + 1)
The last step is obtained from the substitution of (9). where α = F · TSSW /TBI . Here, TBI and TSSW denote
Since ps + p = 1, we can obtain the following equation: the duration of a BI and an SSW frame, respectively. The
 N −1 first term in (14) accounts for the latency caused by i
1 consecutive collisions before a successful transmission.
1− + p − 1 = 0. (11)
M (pR (W − 1) /2 + 1) According to the BF training protocol, if a collision
The conditional collision probability p can be obtained via occurs, the collided STA must wait until the subsequent
solving (11). However, due to the summation and permutation, BI to initiate a transmission attempt, which increases the
Eq. (11) is an implicit function, and thus it is challenging to latency by an entire BI duration. The second term in (14)
obtain a closed-form solution. Here, we apply a numerical represents the latency for the successful transmission. As
method to obtain p. More importantly, Eq. (11) demonstrates shown in Fig. 2, the successful BF training procedure
that the conditional collision probability p depends on retry consists of F (referred to as FSS field in 802.11ad)
limit R, contention window size W , the number of A-BFT SSW frames and the corresponding BF training latency
slots M and the number of STAs N . Obviously, these pa- is F · TSSW . With (14), the average BF training latency
rameters are closely coupled with each other, which poses of an STA when the number of consecutive collisions is
challenges on further BF training efficiency analysis. less than R, is given by
R−1
With the obtained p, the successful transmission probability X
(1 − p) pi E [Di ]
can be computed via the following way. Since 1 − p repre-
i=0
sents the conditional successful transmission probability given R−1
the STA is active, the successful transmission probability is X
= (1 − p) pi (i + α) TBI
represented as follows: i=0
R−1 R−1
!
p̂s = (1 − p) τ (12) X
i
X
i
= TBI (1 − p) p ·i+α p
which also denotes the successful BF training probability. i=0 i=0
pR+1 (R − 1) − RpR + p
 
R


= TBI + 1−p α .
1−p
(15)
C. Average BF Training Latency • When i ≥ R, each collision results in a further random
backoff time of w BIs, and hence E [Di ] is given by
In addition to the average successful BF training probability, E [Di ] = ((i − R + 1)(E [w] + 1) + R − 1)TBI + F · TSSW
the BF training latency is also another important performance
= TBI ((i − R + 1)(E [w] + 1) + R − 1 + α)
indicator for the protocol. The average BF training latency
(16)
represents the average time spent until a successful transmis-
sion. Taking the consecutive collisions before a successful where E [w] denotes the average backoff time. The BF
transmission into account, the average BF training latency can training latency when the number of a consecutive colli-

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
7

sion exceeds R, can be represented by of A-BFT slots and protocol parameters on the BF training
∞
X efficiency.
(1 − p) pi E [Di ] From (19), it is difficult to obtain the explicit relationship
i=R between protocol parameters and the BF training efficiency.
∞
X This is because (19) is a complex function of τ , N , and M .
= TBI (1 − p) pi ((i − R + 1)(E [w] + 1) To make the performance analysis tractable, an approximate
i=R expression of the BF training efficiency is desired. According
+R − 1 + α) to (10), for a large number of STAs, the conditional successful
∞ transmission probability can be approximated by
(a) X
= TBI (1 − p) pR+j ((j + 1) (E [w] + 1)  τ N −1
j=0 p̂s = 1 −
(17) M
+R − 1 + α) (a)  τ N

∞ ≈ 1−
M
(20)
X
= TBI (1 − p) pR (E [w] + 1) pj · j 
τ M/τ
N τ /M
j=0 = 1−
 M
∞
(b)
≈ e−N τ /M
X
+ ((E [w] + 1) + R − 1 + α) pj 
j=0
  where (a) is obtained when N is sufficiently large. Since
(b) R W +1 we consider dense user scenarios, this condition can be eas-
= TBI · p +R+α−1
2(1 − p) ily satisfied, and hence the approximation is valid; and (b)
n
where (a) follows by changing variable j = i − R; (b) follows from the equation lim (1 − 1/n) = 1/e where
n→∞
is due to the substitution of E [w] = (W − 1)/2. n = M/τ is sufficiently large in dense user scenarios.
Taking the above two cases into consideration, the average Here, (20) demonstrates the average successful transmission
BF training latency in (13) can be obtained as follows: probability is dependent on N/M .
Substituting (20) into (19), the approximate BF training
R−1 ∞
X X efficiency when the number of STAs is sufficiently large, can
D= (1 − p) pi E [Di ] + (1 − p) pi E [Di ] be given as follows:
i=0 i=R
pR+1 (R − 1) − RpR + p
  τ N −τ N /M
e Ŝ = . (21)
+ 1 − pR α


= TBI M
1−p (18)

W + 1
 Eq. (21) characterizes the approximate BF training effi-
R
+ TBI p +R+α−1 ciency with respect to system parameters and provides the
2(1 − p)
 R  following important insights for protocol design in dense user
p (W − 1)/2 + p scenarios. Firstly, the BF training efficiency in dense user
= TBI +a .
1−p scenarios depends on the ratio of the number of STAs to the
Eq. (18) demonstrates that the BF training latency depends on number of A-BFT slots, i.e., N/M . Therefore, increasing the
collision probability p, retry limit R and contention window number of A-BFT slots adaptive to the number of STAs is
size W . Obviously, the BF training latency increases with an effective solution to maintain good BF training efficiency.
the increase of the collision probability since severe collisions Secondly, with a further analysis of (21), the BF training
would result in substantial retransmission in the network. efficiency increases with τ N /M when it is less than 1 while
decreases with τ N /M when it exceeds 1. For a practical
system with a fixed number of A-BFT slots, the BF training
V. BF T RAINING E FFICIENCY A NALYSIS efficiency would decrease with the increase of user density
A. Approximate BF Training Efficiency in dense user scenarios. Thirdly, since τ is determined by
the protocol parameters, the BF training efficiency is also
Since the average successful transmission probability is ps τ , affected by the protocol parameters. Thus, we can conclude
the average number of STAs that successfully perform BF that the default protocol parameter setting is not always
training is ps τ N . The BF training efficiency of the protocol optimal in different user density scenarios, which implies that
represents the percentage of A-BFT slots that has been suc- the protocol parameters should be tuned according to user
cessfully utilized, which is defined as density. The accuracy of this approximation is validated by
ps τ N simulation results in Fig. 8.
S=
M (19)
 τ N −1 τ N B. Maximum BF Training Efficiency
= 1−
M M With the approximate BF training efficiency, we target
where the last step follows from the substitution of (10). at analyzing the maximum BF training efficiency in dense
Since τ depends on protocol parameters, above equation user scenarios. The maximum BF training efficiency is the
characterizes the impact of the number of STAs, the number maximum throughput of the BF training protocol, which

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
8

provides theoretical insights on the capacity of the 802.11ad VI. E NHANCEMENT S CHEME
BF training protocol. Since the number of STAs in the network
is uncontrollable, the number of A-BFT slots is optimized with In this section, we propose an enhancement scheme to
respect to the number of STAs to maximize the BF training improve BF training performance in dense user scenarios.
efficiency. Here, we assume the number of A-BFT slots is Previous analysis indicates that the BF training efficiency
sufficient. The problem with the limitation of A-BFT slots depends on the ratio of the number of STAs to the number of
will be discussed in the following section. The BF training A-BFT slots. However, the number of A-BFT slots in current
efficiency maximization problem can be formulated as follows: 802.11ad is limited. Specifically, 802.11ad only provides at
most 8 A-BFT slots for BF training. Even though future
802.11ay increases the number of A-BFT slots to 40, the
P1 :max Ŝ number of A-BFT slots is still limited [34]. The limitation
M (22)
s.t. M ∈ Z+ . of A-BFT slots renders the performance degradation in dense
user scenarios.
The constraint indicates that the number of A-BFT slots To improve the BF training efficiency in dense user net-
takes positive integers. Hence, problem (22) is an integer works, we propose an enhancement scheme that protocol
programming problem. To solve this problem, we first relax parameters (R and W ) should be tuned with the user den-
the integer constraint to a non-integer constraint. Then, this sity in the network. The reasoning behind the enhancement
optimization problem can be readily solved by taking the scheme is that adjusting protocol parameters can determine
derivation of (21). The corresponding condition to achieve the the probability that STAs are active (τ ), such that the number
maximum BF training efficiency is given by of active STAs could be equivalent to the number of provided
M ? = τ N. (23) A-BFT slots. This reason is also validated by our analytical
results. Since τ = 1/(pR (W − 1) /2 + 1) in (9), τ decreases
The condition (23) provides an interesting insight on the with the decrease of R, i.e., a small value of the retry limit
protocol design, i.e., in order to maximize the BF training renders STAs prone to enter backoff states and thus reduces
efficiency, the number of A-BFT slots should equal the number the number of active STAs in the network. In addition, τ
of active STAs (τ N ) in the network. In other words, the also decreases with the decrease of W , because large con-
network should provide equivalent A-BFT slots for all active tention window size makes STAs keep inactive for a longer
users. time. Therefore, adaptively adjusting protocol parameters in
tune with user density is an effective solution in dense user
Under the condition in (23), the maximum BF training
scenarios.
efficiency is
The proposed enhancement scheme mainly consists of two
Ŝ ? = e−1 . (24)
steps. Firstly, AP obtains the number of STAs in the network.
Secondly, based on the number of STAs in the network, AP
Since τ depends on the protocol parameters, the optimal configures the optimal parameter setting of the BF training
number of A-BFT slots is also dependent on the protocol protocol and broadcasts the protocol parameters to all STAs
parameters. When the condition in (23) is satisfied, we have in the network. The key issue of the enhancement scheme is
p = 1 − 1/e because of (20). Then, according to (9), (23) can the optimal protocol parameters, which could be obtained by
be rewritten as solving the following optimization problem with the objective
N to maximize the BF training efficiency:
M? = R
. (25)
−1
(1 − e ) (W − 1) /2 + 1
which characterizes the relationship between the optimal num- P2 :max S
ber of A-BFT slots and protocol parameters (R and W ). With R,W

a further analysis of (25), the optimal number of A-BFT slots s.t. 1 ≤ W ≤ Wmax , W ∈ Z+ (26a)
decreases with the decrease of R and the increase of W . To 1 ≤ R ≤ Rmax , R ∈ Z +
(26b)
achieve the maximum BF training efficiency with limited A-
BFT slots, we should choose a small value of R and a large where Wmax denotes the maximum contention window size,
value of W . The detailed optimization of protocol parameters which is applied to avoid infinite backoff time. Similarly,
is given in Section VI. Rmax denotes the maximum value of the retry limit.
The problem is challenging to be solved due to the following
Remark 2: Approximate analysis of the maximum BF two reasons. Firstly, the optimization variables obey integer
training efficiency reveals two useful insights into the per- constraints. Secondly, the objective function is non-convex.
formance of the BF training protocol. Firstly, the maximum This is because S is an implicit function with respect to R
BF training efficiency is only 1/e. The BF training efficiency and W , since two variables R and W are coupled according
is low because of severe collisions in BF training procedure. to (11). Therefore, problem P2 is an integer non-convex
Secondly, to achieve the maximum BF training efficiency, the problem, which is difficult to obtain the optimal solution.
optimal number of A-BFT slots should equal the number of However, because the number of possible combinations of
active STAs in the network. A mismatch between the active the protocol parameters is limited, we can solve this problem
STAs and A-BFT slots leads to performance degradation. via an exhaustive search method [36]. Since W takes positive

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
9

Table III
100
S IMULATION PARAMETERS [11], [32].

Success BF Training Probability (%)

Sim. M=16
Ana. M=16
Parameter Value 80 Sim. M=12
BI duration (TBI ) 100 ms Ana. M=12
SSW frame duration (TSSW ) 15.8 us Sim. M=8
Number of STAs (N ) [8, 32] 60 Ana. M=8
Number of A-BFT slots (M ) 8
Retry limit (R) 8 40
Contention window size (W ) 8
Number of SSW frames in an A-BFT slot (F ) 16
20 M increases
Frequency band 60 GHz
Maximum retry limit (Rmax ) 20
Maximum contention window size (Wmax ) 20 0
5 10 15 20 25 30
Number of STAs (N)
integer values within (0, Wmax ], the number of possible values
that W can take is Wmax . Similarly, for integer variable Fig. 4. Successful BF training probability in terms of the number of STAs.
R ∈ (0, Rmax ], the number of possible values that R can take
is Rmax . In this way, the number of possible combinations
0.4
of R and W is Rmax Wmax . To obtain the optimal value,

BF Training Efficiency (S)

the exhaustive search method needs to enumerate Rmax Wmax 0.35
combinations. Hence, the computational complexity of the
exhaustive search method is O(Wmax Rmax ). In addition, to 0.3

reduce computational time, the optimal protocol parameters 0.25

Sim. M=16
Ana. M=16
can be computed offline with different numbers of STAs and
Sim. M=12
loaded into AP as a table. Then, based on the number of STAs 0.2
Ana. M=12
in the coverage of AP, the AP could search the table to obtain Sim. M=8
0.15
the optimal protocol parameters with low complexity. Ana. M=8

0.1
5 10 15 20 25 30
VII. S IMULATION R ESULTS Number of STAs (N)
In this section, we evaluate the proposed analytical model
and the enhancement scheme via extensive Monte-Carlo sim- Fig. 5. BF training efficiency with respect to the number of STAs.
ulations.
that the successful BF training probability mainly depends on
A. Simulation Setup the numbers of STAs and A-BFT slots. Note that the analytical
results shown in Figs. 4-7 are obtained via a numerical solution
We validate the proposed analytical model based on a of (11), instead of the approximation in (20) which is valid
discrete event simulator coded in Matlab. We consider an only for a large number of STAs. Hence, the shown results
802.11ad system which operates in the unlicensed 60 GHz are accurate even for a small number of STAs.
band. Specifically, we simulate 10,000 BIs and study the statis-
In Fig. 5, we further show the BF training efficiency with
tics of interests. The BI duration is set to 100 ms according to
respect to the number of STAs. Several important observations
a typical value in practical 802.11ad systems. Unless otherwise
can be made. Firstly, simulation results are closely matched
specified, we set M = 8, W = 8 and R = 8 based on
with analytical results, which further validates our analytical
the default configuration of the 802.11ad standard [11]. Other
model. Secondly, the BF training efficiency exhibits a bell-
important simulation parameters are listed in Table III. For
shape behavior. The reason is two-fold: a) many A-BFT slots
each experiment, we conduct 1000 simulation runs and plot a
are not utilized in low user density scenarios; and b) severe
95 percent confidence interval for each simulation point.
collision occurs in high user density scenarios, which results
in a low BF training efficiency. Thus, to achieve the maximum
B. Analytical Model Validation BF training efficiency, the number of A-BFT slots should be
Figure 4 shows the successful BF training probability in cautiously selected. Thirdly, a system with more A-BFT slots
terms of the number of STAs for M = 8, 12, 16. It can be achieves a higher BF training efficiency than that with fewer
seen that the results obtained via our analytical model are A-BFT slots. For example, in a dense user scenario (N = 32),
highly consistent with that via simulations. As expected, we the BF training efficiency for M = 16 is 25% more than that
can see a lower successful BF training probability in denser for M = 8. This is because more A-BFT slots can effectively
user scenarios since more STAs contend for limited A-BFT alleviate the collision issue in dense user scenarios. Finally, we
slots. The successful BF training probability drops from more can observe the maximum BF training efficiency is around 1/e
than 80% to less than 20% as the number of STAs increases (0.37), which complies with our analytical results.
from 4 to 32. Moreover, as the number of A-BFT slots (M ) Figure 6 shows the impact of the number of STAs on the
increases, the successful BF training probability increases average BF training latency. It can be observed that simulation
because more A-BFT slots are provided. The results validate results comply with our analytical results in (18). Clearly, the

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
10

1.5
BF Training Latency (seconds) 0.4 M=8
Sim. M=8 M=12

BF Training Efficiency (S)

Ana. M=8 M=16
Sim. M=12 0.35 Approx.
1 Ana. M=12
Sim. M=16 M increases
Ana. M=16 0.3

0.328
0.5 0.326
0.25 0.324
0.322
0.32
0.318
0.2 0.316
2.5 2.6 2.7
0
5 10 15 20 25 30 0.5 1 1.5 2 2.5 3
Number of STAs (N) Ratio between # STAs and # A-BFT slots (N/M)

Fig. 6. Average BF training latency with different numbers of STAs. Fig. 8. BF training efficiency with respect to different ratios of the number
of STAs to the number of A-BFT slots.
0.4
BF Training Efficiency (S)

0.35 0.4

BF training efficiency (S)

28%
0.3 Sim. R=8 0.35
Ana. R=8 35%
Sim. R=4
0.25 Ana. R=4 0.3
Sim. R=2
Ana. R=2
0.25 Enhancement Scheme
0.2
5 10 15 20 25 30 Default 802.11ad
Number of STAs (N)
0.2
5 10 15 20 25 30
Fig. 7. BF training efficiency with different values of the retry limit. Number of STAs (N)
(a) M=8
BF training latency increases with the number of STAs. This
is due to a large amount of retransmission of STAs that suffer
from severe collisions in dense user scenarios. Moreover, the 0.4
BF training efficiency (S)

average BF training latency decreases as the number of A-

0.35 17%
BFT slots increases. Specifically, when N = 32, the average
BF training latency for M = 8 is up to 1.3 seconds, which
0.3
is 150% more than that for M = 16. Hence, increasing the
number of A-BFT slots can effectively reduce the BF training 0.25
latency.
Figure 7 plots the impact of the value of the retry limit on Enhancement Scheme
0.2
Default 802.11ad
the BF training efficiency. Specifically, in dense user scenarios,
the protocol with a small value of the retry limit achieves 0.15
5 10 15 20 25 30
a higher BF training efficiency than that with a large value
Number of STAs (N)
of retry limit. For example, when N = 32, the protocol for
(b) M=12
R = 2 achieves around 28% performance gain as compared
to that for R = 8. The reason is that a small value of the Fig. 9. BF training efficiency comparison in terms of different system
retry limit renders STAs susceptible to enter backoff states. As parameters.
such, fewer active STAs will contend for A-BFT slots, thereby
enhancing BF training efficiency in dense user scenarios. More
importantly, since R = 8 is the default configuration of the different numbers of A-BFT slots and STAs are quite close,
retry limit, we claim that the default protocol configuration which complies with our analytical result that the BF training
is not optimal in dense user scenarios. Adaptively adjusting efficiency mainly depends on the ratio of contending STAs to
protocol parameters in tune with the user density is a potential the provided A-BFT slots. In addition, as M increases, the
solution to improve BF training efficiency. gap between simulation results and our approximation in (21)
As plotted in Fig. 8, we show the BF training efficiency in narrows. The gap can be ignored in dense user scenarios, i.e.,
terms of the ratio of the number of STAs to the number of A- the ratio is larger than 2, which validates the accuracy of our
BFT slots. The BF training efficiency with the same ratio for approximation.

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
11

1.5 10
BF Training Latency (seconds)
M=16
Default 802.11ad M=12

Optimal Retry Limit (R*)

Enhancement Scheme 28% 8 M=8
1
6

0.5 4
M increases

2
0
5 10 15 20 25 30 35 5 10 15 20 25 30
Number of STAs (N) Number of STAs (N)
(a) M=8
Fig. 11. The optimal value of the retry limit in terms of the number of STAs.
0.8
probability. As compared to the default 802.11ad, the proposed
BF Training Latency (seconds)

Default 802.11ad 16% scheme can achieve a considerable latency reduction. Specif-
0.6 Enhancement Scheme ically, for a large number of STAs (N = 32), a 28% latency
reduction can be observed clearly in Fig. 10(a). Similar to the
BF training efficiency, as the number of A-BFT slots grows,
0.4
the performance gain on the average BF training latency drops,
because the increase of A-BFT slots relieves the collision. As
0.2 shown in Fig. 10(b), for M = 12, the proposed scheme still
obtains a 16% latency reduction when N = 32, which further
validates the effectiveness of the proposed scheme.
0
5 10 15 20 25 30 35 Finally, Fig. 11 shows the optimal values of the retry limit
Number of STAs (N) with respect to the number of STAs for M = 8, 12, 16. The
(b) M=12 optimal value of the retry limit decreases with the number
of STAs. The simulation results show that a small value
Fig. 10. Average BF training latency comparison in terms of different system
parameters. of the retry limit is preferred in dense user scenarios. For
example, for M = 8, the optimal value of the retry limit
C. Enhancement Scheme Evaluation decreases to 1 when the number of STAs is larger than 28.
The reason is that a small value of the retry limit in dense user
In this subsection, we evaluate the performance of the
scenarios renders STAs prone to enter backoff states. Then, the
proposed enhancement scheme by comparing with the default
collision probability can be reduced to enhance the BF training
802.11ad BF training protocol, in which both R and W adopt
efficiency. In addition, with the increase of A-BFT slots, the
default configurations.
BF training procedure becomes less congested, and hence a
We first evaluate the BF training efficiency of the proposed
larger value of the retry limit should be chosen. For instance,
enhancement scheme in Fig. 9. Several important observations
when N = 32, the optimal value of the retry limit is 3 for
can be obtained from simulation results. Firstly, the proposed
M = 16, which is larger than that for M = 8.
enhancement scheme can significantly enhance the BF training
efficiency in dense user scenarios as compared to the bench-
mark, which is presented in Fig. 9(a). Specifically, we show VIII. C ONCLUSION AND F UTURE W ORK
that a 35% performance gain can be achieved when N = 32. In this paper, we have investigated the performance of the
The key reason is that the proposed scheme can adjust protocol BF training protocol in 802.11ad. An accurate analytical model
parameters in tune with user density to achieve the best has been developed to analyze the BF training efficiency
performance. Secondly, the performance gap between two and latency, which is validated by extensive simulation re-
schemes narrows as the number of A-BFT slots increases. As sults. Theoretical analysis has unveiled the impacts of various
reported in Fig. 9(b) for M = 12, the proposed enhancement protocol components, including the number of A-BFT slots,
scheme only achieves 17% performance gain as compared to user density and the protocol parameters, on the BF training
the default 802.11ad. Hence, simulation results show that the performance. To enhance the BF training performance in dense
proposed enhancement scheme is more suitable for dense user user scenarios, we have proposed an enhancement scheme
scenarios with insufficient A-BFT slots. which can effectively improve the BF training efficiency as
We then present the average BF training latency comparison well as reducing the BF training latency.
between the proposed enhancement scheme and the default Since 802.11ay is expected to adopt a similar BF training
802.11ad protocol for M = 8, 12, as shown in Fig. 10. The protocol, the proposed analytical model can be readily ex-
average BF training latency of two schemes increases with tended to future 802.11ay systems. This work can be viewed
the number of STAs because of the increase of the collision as an effective attempt towards the performance analysis of

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
12

BF training protocol in mmWave networks. For our future [7] N. Kato, Z. M. Fadlullah, B. Mao, F. Tang, O. Akashi, T. Inoue, and
work, we will investigate the protocol performance in mobile K. Mizutani, “The deep learning vision for heterogeneous network traffic
control: Proposal, challenges, and future perspective,” IEEE Wireless
scenarios. Commun., vol. 24, no. 3, pp. 146–153, Jun. 2016.
[8] J. Zhao, F. Gao, Q. Wu, S. Jin, Y. Wu, and W. Jia, “Beam tracking for
IX. ACKNOWLEDGMENT UAV mounted SatCom on-the-move with massive antenna array,” IEEE
J. Sel. Areas in Commun., vol. 36, no. 2, pp. 363–375, Feb. 2018.
This work was financially supported by Huawei Canada [9] W. Peng, W. Gao, and J. Liu, “AI-enabled massive devices multiple
Co., Ltd. The authors thank Yan Xin, Osama Aboul-Magd and access for smart city,” IEEE Internet Things J., vol. 6, no. 5, pp. 7623–
Edward Au from Huawei Canada for valuable suggestions and 7634, Oct. 2019.
[10] L. Zhao, J. Wang, J. Liu, and N. Kato, “Routing for crowd management
comments. The authors thank Qinghua Shen, and Miao Wang in smart cities: A deep reinforcement learning perspective,” IEEE
(Miami University) for very helpful feedback and suggestions Commun. Mag., vol. 57, no. 4, pp. 88–93, Apr. 2019.
on the paper. [11] IEEE Standards Association, “IEEE standards 802.11ad-2012: Enhance-
ment for very high throughput in the 60 GHz band,” 2012.
[12] P. Zhou, K. Cheng, X. Han, X. Fang, Y. Fang, R. He, Y. Long, and
A PPENDIX Y. Liu, “IEEE 802.11ay-based mmWave WLANs: Design challenges
and solutions,” IEEE Commun. Surveys Tuts., vol. 20, no. 3, pp. 1654–
A. Proof of Theorem 1 1681, 3rd Quart. 2018.
Based on the proposed analytical model, the steady state [13] W. Wu, N. Zhang, N. Cheng, Y. Tang, K. Aldubaikhy, and X. Shen,
“Beef up mmWave dense cellular networks with D2D-assisted coop-
probability vector can be solved with the following three steps. erative edge caching,” IEEE Trans. Veh. Technol., vol. 68, no. 4, pp.
Firstly, with the one-step transition probability in (1), we 3890–3904, Apr. 2019.
know that πr+1,0 = p · πr,0 , ∀r ∈ [0, R − 2]. Hence, the steady [14] N. Cheng, F. Lyu, W. Quan, C. Zhou, H. He, W. Shi, and X. Shen,
state probability at state (R − 1, 0) can be represented by “Space/aerial-assisted computing offloading for IoT applications: A
learning-based approach,” IEEE J. Sel. Areas Commun., vol. 37, no. 5,
pp. 1117–1129, May 2019.
πR−1,0 = pR−1 π0,0 . (27)
[15] W. Wu, Q. Shen, M. Wang, and X. Shen, “Performance analysis of
IEEE 802.11.ad downlink hybrid beamforming,” in Proc. IEEE Int. Conf.
Secondly, with the one-step transition probability in backoff Commun., 2017.
states (3)-(5), the steady probability for backoff states can be [16] S. Sur, X. Zhang, P. Ramanathan, and R. Chandra, “BeamSpy: Enabling
given as follows: robust 60 GHz links under blockage,” in Proc. USENIX NSDI, 2016, pp.
193–206.
(W − w) p [17] J. Wang, Z. Lan, C. Pyo, T. Baykas, C. Sum, M. A. Rahman, J. Gao,
πR,w = (πR,0 + πR−1,0 ) , ∀w ∈ [0, W − 1]. (28) R. Funada, F. Kojima, H. Harada, and S. Kato, “Beam codebook based
W
beamforming protocol for multi-Gbps millimeter-wave WPAN systems,”
p
Specifically, it can be obtained that πR,0 = 1−p πR−1,0 by IEEE J. Sel. Areas Commun., vol. 27, no. 8, pp. 1390–1399, Oct. 2009.
taking w = 0 in (28). Thus, according to (27), (28) can be [18] Z. Marzi, D. Ramasamy, and U. Madhow, “Compressive channel esti-
mation and tracking for large arrays in mm-Wave picocells,” IEEE J.
rewritten as Sel. Topics Signal Process., vol. 10, no. 3, pp. 514–527, Apr. 2016.
(W − w) p [19] S. Sur, I. Pefkianakis, X. Zhang, and K. H. Kim, “WiFi-assisted 60 GHz
πR,w = πR−1,0 wireless networks,” in Proc. ACM MOBICOM, 2017, pp. 28–41.
W (1 − p) [20] Y. Zhou, Z. M. Fadlullah, B. Mao, and N. Kato, “A deep-learning-based
(29)
(W − w) pR radio resource assignment technique for 5G ultra dense networks,” IEEE
= π0,0 , ∀w ∈ [0, W − 1]. Network, vol. 32, no. 6, pp. 28–34, Nov./Dec. 2018.
W (1 − p) [21] Y.-J. Yu, T.-Y. Hsieh, and A.-C. Pang, “Millimeter-wave backhaul traffic
Finally, substituting (27) and (29) into (7b), π0,0 can be minimization for CoMP over 5G cellular networks,” IEEE Trans. Veh.
Technol., vol. 68, no. 4, pp. 4003–4015, Apr. 2019.
solved as [22] Y. Wang, D. Yu, Q. Liu, and F. C. Lau, “Inductive coloring: Implement-
(1 − p) ing basic communication primitives with Rayleigh-fading interference,”
π0,0 = R . (30)
p (W − 1) /2 + 1 in Proc. IEEE INFOCOM, 2016, pp. 1–9.
[23] M. Takahashi, Y. Kawamoto, N. Kato, A. Miura, and M. Toyoshima,
Since other steady state probabilities can be represented by “Adaptive power resource allocation with multi-beam directivity control
π0,0 , as shown in (27) and (29), Theorem 1 is proved. in high-throughput satellite communication systems,” IEEE Wireless
Commun. Lett., vol. 8, no. 4, pp. 1248–1251, Aug. 2019.
[24] M. Hashemi, C. E. Koksal, and N. B. Shroff, “Out-of-band millimeter
R EFERENCES wave beamforming and communications to achieve low latency and high
[1] O. Abari, D. Bharadia, A. Duffield, and D. Katabi, “Enabling high- energy efficiency in 5G systems,” IEEE Trans. Commun., vol. 66, no. 2,
quality untethered virtual reality,” in Proc. USENIX NSDI, 2017, pp. pp. 875–888, Feb. 2018.
531–544. [25] H. Hassanieh, O. Abari, M. Rodriguez, M. Abdelghany, D. Katabi,
[2] J. Wang, L. Zhao, J. Liu, and N. Kato, “Smart resource allocation for and P. Indyk, “Fast millimeter wave beam alignment,” in Proc. ACM
mobile edge computing: A deep reinforcement learning approach,” IEEE SIGCOMM, 2018, pp. 432–445.
Trans. Emerg. Topics Comput., doi: 10.1109/TETC.2019.2902661, 2019. [26] W. Wu, N. Cheng, N. Zhang, P. Yang, W. Zhuang, and X. Shen, “Fast
[3] K. Ota, M. Dong, J. Gui, and A. Liu, “QUOIN: Incentive mechanisms mmwave beam alignment via correlated bandit learning,” IEEE Trans.
for crowd sensing networks,” IEEE Network, vol. 32, no. 2, pp. 114–119, Wireless Commun., vol. 18, no. 12, pp. 5894–5908, Dec. 2019.
Mar./Apr.2018. [27] P. Zhou, X. Fang, Y. Fang, R. He, Y. Long, and G. Huang, “Beam
[4] H. Li, K. Ota, M. Dong, and M. Guo, “Learning human activities management and self-healing for mmwave UAV mesh networks,” IEEE
through Wi-Fi channel state information with multiple access points,” Trans. Veh. Techol., vol. 68, no. 2, pp. 1718–1732, Feb. 2019.
IEEE Commun. Mag., vol. 56, no. 5, pp. 124–129, May 2018. [28] P. Zhou, X. Fang, Y. Fang, Y. Long, R. He, and X. Han, “Enhanced
[5] H. Li, K. Ota, and M. Dong, “ECCN: Orchestration of edge-centric random access and beam training for millimeter wave wireless local
computing and content-centric networking in the 5G radio access networks with high user density,” IEEE Trans. Wireless Commun.,
network,” IEEE Wireless Commun., vol. 25, no. 3, pp. 88–93, Jun. 2018. vol. 16, no. 12, pp. 7760–7773, Dec. 2017.
[6] X. Shen, J. Gao, W. Wu, K. Lyu, M. Li, W. Zhuang, X. Li, and J. Rao, [29] S. Shao, H. Zhang, D. Koutsonikolas, and A. Khreishah, “Two-
“AI-assisted network-slicing based next-generation wireless networks,” dimensional reduction of beam training overhead in crowded 802.11ad
IEEE Open Journal of Vehicular Technology, vol. 1, no. 1, pp. 45–66, based networks,” in Proc. IEEE INFOCOM Workshops, 2018, pp. 680–
2020. 685.

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2020.2982178, IEEE
Transactions on Vehicular Technology
13

[30] S. G. Kim, K. Jo, S. Park, H. Cho, J. Kim, and S. Bang, “On random Ning Zhang (M’15-SM’18) received the Ph.D de-
access in A-BFT,” IEEE 802.11 Documents, doc.:IEEE 802.11-16/0948- gree from University of Waterloo, Canada, in 2015.
00-00ay, Jul. 2016. After that, he was a postdoc research fellow at
[31] Y. Xin, R. Sun, and O. Aboul-Magd, “Channel access in A-BFT University of Waterloo and University of Toronto,
over multiple channels,” IEEE 802.11 Documents, doc.:IEEE 802.11- Canada, respectively. Since 2017, he has been an As-
16/0101r0, Jan. 2016. sistant Professor at Texas A&M University-Corpus
[32] K. Jo, S. Park, H. Cho, J. Kim, S. Bang, and S. G. Kim, “Short SSW Christi, USA. He serves as an Associate Editor of
frame for A-BFT,” IEEE 802.11 Documents, doc.:IEEE 802.11-17/0117- IEEE Internet of Things Journal, IEEE Transac-
00-00ay, Jan. 2017. tions on Cognitive Communications and Network-
[33] K. Aldubaikhy, W. Wu, Q. Ye, and X. Shen, “Low-complexity user se- ing, IEEE Access and IET Communications. His
lection algorithm for multiuser transmission in mmwave WLANs,” IEEE current research interests include wireless communi-
Trans. Wireless Commun., doi: 10.1109/TWC.2020.2964551, 2020. cation and networking, mobile edge computing, machine learning and physical
[34] Y. Ghasempour, C. R. da Silva, C. Cordeiro, and E. W. Knightly, “IEEE layer security.
802.11ay: Next-generation 60 GHz communication for 100 Gb/s Wi-Fi,”
IEEE Commun. Mag., vol. 55, no. 12, pp. 186–192, Dec. 2017.
[35] D.-Y. Yang, T.-J. Lee, K. Jang, J.-B. Chang, and S. Choi, “Performance Peng Yang (S’16-M’18) received his B.E. degree
enhancement of multirate IEEE 802.11 WLANs with geographically in Communication Engineering and Ph.D. degree in
scattered stations,” IEEE Trans. Mobile Comput., vol. 5, no. 7, pp. 906– Information and Communication Engineering from
919, Jul. 2006. Huazhong University of Science and Technology
[36] T. H. Luan, X. Ling, and X. Shen, “MAC in motion: Impact of mobility (HUST), Wuhan, China, in 2013 and 2018, respec-
on the MAC of drive-thru Internet,” IEEE Trans. Mobile Comput., tively. He was with the Department of Electrical
vol. 11, no. 2, pp. 305–319, Feb. 2012. and Computer Engineering, University of Waterloo,
[37] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed Canada, as a Visiting Ph.D. Student from Sept. 2015
coordination function,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, to Sept. 2017, and a Postdoctoral Fellow from Sept.
pp. 535–547, Mar. 2000. 2018 to Dec. 2019. Since Jan. 2020, he has been
[38] Y. Ren, L. Li, G. Xie, Y. Yan, Y. Cao, H. Huang, N. Ahmed, Z. Zhao, a faculty member with the School of Electronic
P. Liao, C. Zhang, G. Caire, A. F. Molisch, M. Tur, and A. E. Willner, Information and Communications, HUST. His current research focuses on
“Line-of-sight millimeter-wave communications using orbital angular mobile edge computing, video streaming and analytics.
momentum multiplexing combined with conventional spatial multiplex-
ing,” IEEE Trans. Wireless Commun., vol. 16, no. 5, pp. 3151–3161,
May 2017.
[39] Y. Zheng, K. Lu, D. Wu, and Y. Fang, “Performance analysis of IEEE
802.11 DCF in imperfect channels,” IEEE Trans. Veh. Technol., vol. 55,
Khalid Aldubaikhy (S’10) is currently an Assis-
no. 5, pp. 1648–1656, Sep. 2006.
tant Professor at Qassim University, Buraydah, Al-
[40] Q. Ye and W. Zhuang, “Distributed and adaptive medium access control
Qassim, Saudi Arabia. He received the B.E. degree
for Internet-of-Things-enabled mobile networks,” IEEE Internet Things
from Qassim University, Saudi Arabia, in 2008,
J., vol. 4, no. 2, pp. 446–460, Apr. 2017.
the M.A.Sc. degree in Electrical and Computer En-
[41] Q. Ye, W. Zhuang, L. Li, and P. Vigneron, “Traffic-load-adaptive
gineering from Dalhousie University, Halifax, NS,
medium access control for fully connected mobile ad hoc networks,”
Canada, in 2012, and the Ph.D. degree in Electrical
IEEE Trans. Veh. Technol., vol. 65, no. 11, pp. 9358–9371, Nov. 2016.
and Computer Engineering from University of Wa-
terloo, Waterloo, ON, Canada, in 2019. His research
interests include millimeter-wave wireless networks,
medium access control, and mmwave 5G networks.

Wen Wu (S’13-M’20) earned the Ph.D. degree in

Electrical and Computer Engineering from Univer- Xuemin (Sherman) Shen (M’97-SM’02-F’09) re-
sity of Waterloo, Waterloo, ON, Canada, in 2019. ceived the Ph.D. degree in electrical engineering
He received the B.E. degree in Information Engi- from Rutgers University, New Brunswick, NJ, USA,
neering from South China University of Technology, in 1990. He is currently a University Professor
Guangzhou, China, and the M.E. degree in Electrical with the Department of Electrical and Computer
Engineering from University of Science and Tech- Engineering, University of Waterloo, Canada. His
nology of China, Hefei, China, in 2012 and 2015, research focuses on network resource management,
respectively. Starting from 2019, he works as a Post- wireless network security, Internet of Things, 5G and
doctoral fellow with the Department of Electrical beyond, and vehicular ad hoc and sensor networks.
and Computer Engineering, University of Waterloo. He is a registered Professional Engineer of Ontario,
His research interests include millimeter-wave networks and AI-empowered Canada, an Engineering Institute of Canada Fellow, a
wireless networks. Canadian Academy of Engineering Fellow, a Royal Society of Canada Fellow,
a Chinese Academy of Engineering Foreign Fellow, and a Distinguished
Lecturer of the IEEE Vehicular Technology Society and Communications
Society.
Dr. Shen received the R.A. Fessenden Award in 2019 from IEEE, Canada,
James Evans Avant Garde Award in 2018 from the IEEE Vehicular Technology
Nan Cheng (M’16) received the Ph.D. degree from Society, Joseph LoCicero Award in 2015 and Education Award in 2017
the Department of Electrical and Computer Engi- from the IEEE Communications Society. He has also received the Excellent
neering, University of Waterloo in 2016, and B.E. Graduate Supervision Award in 2006 and Outstanding Performance Award 5
degree and the M.S. degree from the Department times from the University of Waterloo and the Premier’s Research Excellence
of Electronics and Information Engineering, Tongji Award (PREA) in 2003 from the Province of Ontario, Canada. He served as
University, Shanghai, China, in 2009 and 2012, the Technical Program Committee Chair/Co-Chair for the IEEE Globecom’16,
respectively. He is currently a professor with School the IEEE Infocom’14, the IEEE VTC’10 Fall, the IEEE Globecom’07,
of Telecommunication Engineering, and with State the Symposia Chair for the IEEE ICC’10, and the Chair for the IEEE
Key Lab of ISN, Xidian University, Shaanxi, China. Communications Society Technical Committee on Wireless Communications.
He worked as a Post-doctoral fellow with the De- He was the Editor-in-Chief of the IEEE INTERNET OF THINGS JOURNAL
partment of Electrical and Computer Engineering, and IEEE Network, and the Vice President on Publications of the IEEE
University of Toronto, from 2017 to 2018. His current research focuses on Communications Society.
space-air-ground integrated system, big data in vehicular networks, and self-
driving system.

0018-9545 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 27,2020 at 18:11:10 UTC from IEEE Xplore. Restrictions apply.