You are on page 1of 15

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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 1

Resource Management for Millimeter-Wave


Ultra-Reliable and Low-Latency Communications
Rui Liu, Guanding Yu, Senior Member, IEEE, Jiantao Yuan, and Geoffrey Ye Li, Fellow, IEEE

Abstract—Many mission-critical and latency-sensitive appli- rate of less than 10−5 for a single transmission of 32-byte long
cations require ultra-reliable and low-latency communications packet [2]. Such rigorous requirements have propelled novel
(URLLC), which has been listed as a new service category of 5G designs in the physical layer, including more conservative
New Radio (NR). To guarantee stringent latency and reliability
constraints, URLLC services always exclusively occupy the spec- coding and modulation methods, non-square packet structure,
trum and have priority over enhanced mobile broadband (eMBB) mini-slot scheduling, and scalable numerology of subcarrier
communications in the current coexistence scenario, which will spacing [3].
greatly affect the performance of eMBB services and degrade On the other hand, medium access control (MAC) layer
the utilization efficiency of the spectrum resource. On the other design is a particularly challenging task when considering the
hand, millimeter-wave (mmWave) communications can fulfill the
enormous throughput requirements of 5G cellular communi- coexistence among diverse 5G use cases, especially between
cations. In this paper, we introduce mmWave communications URLLC and ongoing enhanced mobile broadband (eMBB).
into URLLC systems to provide a more efficient coexistence The original concept generally assumes an orthogonal ap-
for eMBB and URLLC. A novel mmWave URLLC system is proach to mitigate the inter-service interference, where the
first developed, where URLLC users are allowed to share the spectrum resources of one service are isolated from the
spectrum resources with eMBB users. Besides, multi-connectivity
technology, which enables users to access multiple base stations others. Under this circumstance, URLLC traffic is always
simultaneously, is introduced to the mmWave URLLC system to scheduled on top of eMBB transmissions due to its stringent
enhance the reliability. Then, a resource management problem requirements. A semi-static scheduling mechanism has been
is formulated, which maximizes the throughput of eMBB users developed [4], where URLLC traffic is allowed to overlay onto
while guaranteeing the latency and reliability requirements of previously allocated eMBB transmissions for immediately
URLLC users. To obtain optimal solutions, we first divide it
into three subproblems, i.e., power allocation, resource matching, transmitting. Besides, to further reduce the latency of URLLC,
and user paring, and then solve them respectively. Simulation a certain amount of time-frequency resources will be reserved
results demonstrate the data rate improvement compared against for URLLC traffic even if the activity of URLLC is low
the traditional coexistence scenario without reusing strategy. [5]. These orthogonal scheduling-based approaches obviously
Moreover, the multi-connectivity functionality poses a great effect degrade the performance of eMBB services and reduce the
on guaranteeing the latency and reliability requirements for
URLLC users. utilization efficiency of the valuable time-frequency resources.
This phenomenon motivates us to develop a new scheduling
Index Terms—Millimeter-wave communications, URLLC, re-
strategy based on non-orthogonal spectrum sharing between
source management, multi-connectivity.
URLLC and eMBB services.
Confronted with the looming spectrum crunch, millimeter-
I. I NTRODUCTION wave (mmWave) communications technology, which extends
Ultra-reliable and low-latency communications, also known cellular communications to the higher mmWave frequency
as URLLC, is a new service category supported by 5G bands, has demonstrated a great potential to achieve im-
New Radio (NR), catering to various types of mission-critical mense network capacity requirements of eMBB services in
use cases, such as factory automation, remote surgery, and 5G [6]. However, few studies have paid attention to the
autonomous driving [1]. As stipulated by the 3rd Generation applications of mmWave communications in URLLC. In fact,
Partnership Project (3GPP) Release 15, URLLC has a target high frequency causes higher phase noise, for which a larger
user-plane radio latency of sub-milliseconds with a packet drop subcarrier spacing is needed to mitigate its impact. Based
on scalable numerology, a shorter transmission time interval
Manuscript received June 12, 2020; revised September 29, 2020; accepted
October 27, 2020. This work was supported by the National Natural Science can be achieved with a larger subcarrier spacing. Therefore,
Foundation Program of China under Grant 61671407 and by the Fundamental mmWave communications are intrinsically suitable for realiz-
Research Funds for the Central Universities. The associate editor coordinating ing URLLC by dint of its high operating frequency. Besides,
the review of this paper and approving it for publication was D. Niyato.
(Corresponding author: Jiantao Yuan.) two characteristics of mmWave communications can well
Rui Liu and Guanding Yu are with the Zhejiang Provincial Key Laboratory support URLLC. The first one is beamforming. Owing to the
of Information Processing, Communication and Networking, Zhejiang Uni- short wavelengths, the elaborate radio frequency chips can
versity, Hangzhou 310027, China (e-mail: {rliu, yuguanding}@zju.edu.cn).
Jiantao Yuan is with the Institute of Ocean Sensing and Networking, be packed into a relatively small physical dimension and a
Ocean College, Zhejiang University, Zhoushan 316021, China (email: yuan- large number of antenna elements can be implemented on the
jiantao@zju.edu.cn). user equipment, forming very narrow beams. The consider-
Geoffrey Ye Li is with the Department of Electrical and Electronic
Engineering, Imperial College London, London SW7 2AZ, U.K. (e-mail: able directivity gain can improve the coverage of mmWave
geoffrey.li@imperial.ac.uk). transmission. Moreover, the spatial freedom provided by a

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 2

large number of antenna elements can mitigate co-channel resource management problem can be achieved.
interference when eMBB and URLLC users share the same •Simulation results validate the effectiveness of our pro-
spectrum. The second one is multi-connectivity. Transmissions posed coexistence scenario and resource management
over mmWave frequencies may experience severe penetra- algorithm. We also summarize several design guidelines
tion attenuation when encountering solid obstacles, for which under common environment parameters.
the performance and robustness of mmWave communications The rest of this paper is organized as follows. The related
will significantly deteriorate. To cope with it, one effective works are reviewed in Section II. The system model is shown
technique is the multi-connectivity, which enables users to in Section III, where the resource management problem is
be associated with multiple base stations (BSs) simultane- also formulated. Then, we introduce an optimal algorithm in
ously. With multi-connectivity, the arriving packets of the user Section IV, where a suboptimal algorithm is also developed to
equipment can be transmitted through multiple transmission reduce the computational complexity. Section V validates the
paths. As a result, the packet drop rate due to buffer over- effectiveness of the proposed scenario and algorithms. Finally,
flow and time constraints can be prominently reduced. In a the whole paper is concluded in Section VI.
nutshell, benefiting from the properties of beamforming and
multi-connectivity, mmWave URLLC has a great potential to
guarantee the strict latency and reliability requirements. II. R ELATED W ORKS
This paper aims to develop a novel non-orthogonal coex- In this section, we will summarize the existing studies about
istence scheme for eMBB and URLLC users to guarantee resource allocation for the URLLC and mmWave communi-
fairness between the two air interfaces and improve the cations, respectively.
spectrum utilization efficiency, where URLLC users reuse the
spectrum of eMBB users. To mitigate the effect of inter- A. Resource Allocation for URLLC
service interference, we introduce mmWave communications
into URLLC services, leveraging beamforming technique to Resource allocation for URLLC has recently attracted a
achieve considerable directivity gain. Moreover, to conquer lot of attention. A general resource allocation problem for
the instability of mmWave communications, we adopt multi- typical URLLC systems has been investigated in [7], where
connectivity functionality to build multiple simultaneous trans- the properties of the power allocation constraints have been
missions for users. Meanwhile, resource management plays analyzed. By dint of it, the global optimal power and band-
a dominant effect on system performance. Thus, this paper width allocation can be achieved through well-established
also develops efficient algorithms for the optimal resource methods. To save the bandwidth resource [8], the bandwidth
management and provides design guidelines for URLLC- is not assigned to inactive uplink users and the packets to
eMBB coexistence scenarios. In brief, the main contributions different users are buffered in one queue at the BS. Based
of this paper are summarized as follows. on this, a joint downlink and uplink bandwidth allocation
algorithm has been provided. To tackle the fast-varying in-
• We introduce mmWave communications into URLLC stantaneous channel state information (CSI) and road-traffic
services. Leveraging the beamforming technique to mit- fluctuation in vehicular networks, a two-stage twin-timescale
igate the effect of co-channel inter-service interference, resource allocation scheme has been developed in [9], where
a novel non-orthogonal coexistence scenario is proposed, the bandwidth is assigned on a long-term timescale but the
where URLLC users can reuse the spectrum resources power is adjusted on a short-term timescale. After analyz-
of eMBB users to access mmWave BSs. Moreover, the ing the impact of hybrid automatic repeat request (HARQ)
URLLC queueing system with two key characteristics schemes on URLLC, a bandwidth allocation algorithm for
of mmWave communications, i.e., multi-connectivity and the repetition code-based packet re-transmission scheme has
channel blockage, is analyzed for the first time. been developed in [10]. A novel resource allocation scheme
• We formulate a resource management problem to maxi- for retransmissions with low resource consumption has been
mize the throughput of eMBB users while satisfying the considered in [11] to satisfy the requirements of the under-
latency and reliability constraints of URLLC users. lying service and the traffic characteristics. A prediction and
• To solve the challenging problem, we equivalently de- communication co-design framework has been developed in
compose it into three subproblems, i.e., the power allo- [12] to further reduce the end-to-end delay, where the wire-
cation problem, the resource matching problem, and the less resources and prediction horizon are jointly optimized.
user pairing problem. For the power allocation problem, Besides, non-orthogonal multiple access (NOMA) has been
we first propose a criterion to judge whether a feasible introduced to the URLLC system. The downlink multiple-
solution exists. Then, the optimal power allocation is input multiple-output (MIMO)-NOMA URLLC system has
developed. For the resource matching problem, we first been investigated in [13], where power allocation algorithms
develop an optimal algorithm based on the minimum cost have been developed. Meanwhile, it has been shown in [14]
network flow algorithm. Additionally, a suboptimal algo- that the NOMA scheme outperforms orthogonal multiple ac-
rithm with low computational complexity is proposed, cess (OMA) scheme when the blocklength is very limited.
which is based on the matching theory. Afterward, the However, the above works have not integrated the mmWave
user pairing problem is solved through the Hungarian communications to the URLLC system, thus the impact of the
algorithm. Based on these, the solutions of the proposed mmWave communication characteristics on URLLC, including

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 3

the multi-connectivity and the channel blockage, has not been


User Paring
investigated in the literature.
Some recent works have explored the coexistence among
eMBB User
URLLC and other cases in 5G, where the network slicing
technique is enabled to ensure that the URLLC service will not
be squeezed from software and hardware resources. A service
multiplexing mechanism for the eMBB and URLLC has been
developed in [15], where multicast transmission and unicast
transmission are adopted by the eMBB slice and the URLLC MmWave BS
slice, respectively. The puncturing scheme has been studied in
[16] and [17], where the arriving URLLC users are scheduled
URLLC user
immediately while the performance of overlapped eMBB users
is sacrificed. Note that the inter-service interference isolation
is assumed in all the aforementioned studies, which is not eMBB Link
URLLC Link
spectrum-efficient. Moreover, as we have mentioned before,
Interference Link
the puncturing scheme is unfair to eMBB users. Therefore, in
this paper, we consider a more efficient coexistence scenario Fig. 1: System model.
for eMBB and URLLC, where URLLC users are allowed to
share the spectrum resources with eMBB users. Only a few studies have focused on resource allocation for
mmWave URLLC systems. A joint power allocation and user
B. Resource Allocation for mmWave Communications scheduling algorithm has been developed in [32] for multi-
hop multi-path self-backhauled mmWave URLLC systems.
There have also been several works on the resource The integration of mmWave communications and massive
allocation for mmWave networks. A Lagrangian dual MIMO in [33] simultaneously provides an order of magnitude
decomposition-based algorithm in [18] optimizes the energy capacity improvement and latency reduction by extending the
efficiency of the mmWave networks, where mmWave BSs Lyapunov framework to incorporate probabilistic latency and
with energy harvesting are ultra-densely deployed. The joint reliability constraints. However, two key characteristics of
user association and resource allocation of the multi-band mmWave communications, i.e., the multi-connectivity and the
mmWave heterogeneous network has been studied in [19], channel blockage, are neither considered in these studies. To
where users can access multiple mmWave frequency bands of the best of our knowledge, our work is the first to analyze
the associated mmWave BS. A cluster-based user association the queueing performance and develop resource management
and resource allocation scheme has been investigated in [20], algorithms for mmWave URLLC with the consideration of
where the femto access points and femto users are clustered. both multi-connectivity and channel blockage.
The spectrum sharing scenario has been studied in [21]–
[23], where the cell association, interference coordination, and III. S YSTEM M ODEL AND P ROBLEM F ORMULATION
beamforming design are jointly considered. The tremendous In this section, we first present the system model, trans-
available bandwidth of mmWave frequency band has also mission model, and queueing model of the proposed non-
been leveraged to implement the backhaul connections [24] orthogonal coexistence scenario. Then, a resource management
and the fronthaul connections [25]. In [24], a joint discrete problem is formulated.
power control and non-unified transmission duration allocation
algorithm has been proposed for a self-backhauling system. A. System Model
In [25], an ultra-dense-cloud radio access network (CRAN) We consider a mmWave network with N URLLC users in
architecture jointly optimizes resource allocation over the a set U = {U1 , U2 , · · · , Un , · · · , UN }, M eMBB users in a
mmWave fronthaul and wireless transmission. set E = {E1 , E2 , · · · , Em , · · · , EM }, and S mmWave BSs
To further enhance the spectrum efficiency, NOMA has been in a set B = {B1 , B2 , · · · , Bs , · · · , BS }, as shown in Fig. 1.
applied in mmWave communications. The joint beamforming Without loss of generality, it is assumed that both users and
and power allocation algorithm in [26] achieves the optimal BSs are uniformly located. We further assume that N < M
max-min fairness for a downlink cellular system with analog since eMBB services are more common than URLLC services
beamforming. To cover the shaded area of mmWave com- in practical scenarios.
munications and address the blockage issue, a sub-channel The multi-connectivity functionality is enabled on both
allocation scheme has been developed in [27] for full-duplex eMBB and URLLC users. Each of them is equipped with
(FD) mmWave device-to-device (D2D) communications. Re- multiple A/D or D/A conversion paths and thus can sup-
source allocation algorithms have been developed in [28] for port multiple mmWave communication links simultaneously.
relay-assisted mmWave networks and in [29] for the conceived Meanwhile, each mmWave BS can serve multiple users at the
mmWave mobile edge computing (MEC). Our previous works same time, where the orthogonal frequency division multiple
in [30], [31] have utilized machine learning techniques for access (OFDMA) is adopted.
the user association strategies where the multi-connectivity is The whole spectrum resource is divided into several or-
enabled. thogonal channels with bandwidth of b. Each channel can

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 4

jnsu TABLE I: Notations in Fig. 1.


z nsu Notation Parameter
φbsn mmWave BS Bs URLLC user Un
φbkm Boresight mmWave BS Bk eMBB user Em
URLLC user U n φuns angle from URLLC user Un mmWave BS Bs
j snb φemk eMBB user Em mmWave BS Bk
z nku b
ζsn mmWave BS Bs URLLC user Un
z snb z knb b
ζsm mmWave BS Bs eMBB user Em
b to
ζkn mmWave BS Bk URLLC user Un
b
z sm b b
ζkm Geographical mmWave BS Bk eMBB user Em
jkm b u
mmWave BS z km ζns angle from URLLC user Un mmWave BS Bs
e  





u
Bs z ms e
jmk
ζnk URLLC user Un mmWave BS Bk
e
ζms eMBB user Em mmWave BS Bs
e mmWave BS e
z mk ζmk eMBB user Em mmWave BS Bk


 Bk
eMBB user Em
e b
Similarly, we define gmk,km and gkm,mk as the transmission
Fig. 2: Transmission model in the scenario of ρmk
ns = 1. directivity gain and reception directivity gain between the
beam of Em directed to Bk and the beam of Bk directed
u b
be utilized by mmWave BSs to serve eMBB users. It is to Em , respectively. According to [34], gns,sn and gsn,ns can
assumed that each eMBB user has already selected several be expressed as
 θnu
channels to access one mmWave BS before we investigate the 
 ∆, if < |φuns − ζns
u
|


resource management of URLLC users. Denote Am as the set  2
θ u
of mmWave BSs that Em is associated with. To avoid co- u
gns,sn (θnu , φuns , ζns
u
)= < 2π − n ,

 2

 2π − (2π − θn ) ∆ , otherwise,
channel interference among eMBB users, it is also assumed 
u
that each channel can only be occupied by one eMBB user. θnu
Since URLLC services are generally instantaneous, tran- (1)

sitory, and intermittent, no spectrum resource is reserved
 θsb

 ∆, if < φbsn − ζsn
b
for URLLC users exclusively. Therefore, in our proposed 
 2
coexistence scenario, URLLC users will reuse the spectrum ( b b )  θb
b b
gsn,ns θs , φsn , ζsn = < 2π − s ,
resources of eMBB users for transmission. Since interference 
 ( ) 2

 2π − 2π − θsb ∆
can be better handled in uplink rather than in downlink, we 
 , otherwise,
assume that URLLC users reuse the uplink channels of eMBB θsb
users. To reduce the amount of information exchange between (2)
URLLC and eMBB users, we assume that each URLLC user where θnu and θsb is the beamwidth of Un and Bs respectively,
can reuse the uplink bandwidth of only one eMBB user, and ∆ ≪ 1. Similarly, we can easily obtain the expressions
e b
and each eMBB user can share its spectrum with only one of gmk,km and gkm,mk in the same way.
URLLC user. This can guarantee that all URLLC users can Then, the signal-to-interference-plus-noise ratio (SINR) of
be served since N < M . Owing to the enormous amount Bs receiving from Un can be expressed as
of available spectrum in mmWave frequencies and the multi- u
puns gns,sn
u c b
gns gsn,ns
ξns = . (3)
connectivity technique, reusing spectrum resources of only one |A |
∑ ∑ mk e e
M m
bN0 + c b
ρns pmk gmk,sn gms gsn,km
eMBB user is enough to guarantee the latency and reliability
m=1 k=1
requirements of URLLC users, which will be further verified
by the simulation results in Section V. where puns and pemk represent the transmit powers from Un to
Since each eMBB user transmits on multiple mmWave links Bs and from Em to Bk , respectively. Accordingly, the SINR
simultaneously, one URLLC user can select and reuse one or of Bk receiving from Em can be expressed as
more uplink channels at the same time to improve reliability e
pemk gmk,km
e c
gmk b
gkm,mk
ξmk = . (4)
and reduce the latency. Besides, each URLLC user also needs ∑
N ∑ S
mk u u c b
to decide which mmWave BSs it will access. Let the binary bN0 + ρns pns gns,km gnk gkm,ns
n=1 s=1
variable ρmk
ns represent the joint user pairing and resource
matching indicator, i.e., ρmk
ns = 1 represents that Un reuses
an uplink channel to access Bs and the channel has already C. Queueing Model
been allocated to Em to access Bk .
Several important factors should be considered for the
mmWave URLLC system, including the multi-connectivity,
B. Transmission Model the rigorous time constraint of each packet, and the channel
We adopt the transmission model shown in Fig. 2, where blockage of mmWave channels. As illustrated in [35], a buffer
the definitions of the notations are summarized in Table I. Let is equipped in each URLLC user to cache its arriving packets,
u b
gns,sn and gsn,ns denote the transmission directivity gain and whose size is finite due to the hardware limitation. Therefore,
reception directivity gain between the beam of Un directed in this paper, we adopt the M/M/Cn /D queueing with packet
to Bs and the beam of Bs directed to Un , respectively. reneging and channel blockage model to analyze the queueing

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 5

system of each mmWave URLLC user. We take Un as an equal to one. Based on this, we can obtain the steady-state
example to briefly introduce the queueing model. probability vectors by
Compared with the M/M/1 queueing system that has been  

c+1
widely used in the literature, our proposed queueing system Pnc = PnCn  ϕnj+1  , 0 ≤ c ≤ Cn − 1. (8)
has several distinctions. First, there are Cn serving channels j=1
for Un . Secondly, a finite buffer of size D is considered
for Un . When the buffer of Un is full, the upcoming packet Based on the steady-state probabilities, we can derive sev-
will be dropped, which affects the reliability of the URLLC eral important performance measures of the queueing system,
transmission. Thirdly, a packet may be dropped if it has not listed in the following.
been transmitted during a period of time tw . This time is a • The expected number of packets in the queue of Un can
random variable following a negative exponential distribution be given by
1
with mean . It should be noted that, tw is just the bound of ∑ ∑n
Cn D+C
w
the queueing delay. As discussed in [36], delay components Lnq = [d − (Cn − c)] pnc,d . (9)
may include transmission delay, queueing delay, processing c=0 d=Cn −c
delay in the radio access network, and backhaul delay. In • Since the waiting time, tw , is a random variable following
the proposed system, the backhaul delay is negligible and the 1
processing delay can also be ignored since the proposed algo- a negative exponential distribution with mean and
w
rithm in this paper has polynomial computational complexity the packets are independent of each other, the instan-
(see Section IV). As discussed in [7], the transmission delay, taneous rate of loss packets due to the time constraint
including pilot transmission for channel estimation and CSI is [d − (Cn − c)] w. Therefore, the average rate of loss
feedback, can be finished within 0.1 ms, which can also be packets due to the time constraint in the queue of Un can
treated as a constant. As a comparison, the queueing delay be given by
could be several times longer, predominantly depending on ∑
Cn ∑n
D+C
the system model and environment parameters. Moreover, in Lntime = [d − (Cn − c)]wpnc,d . (10)
our model, the queueing delay is relevant to reliability. Thus, c=0 d=Cn −c+1
we mainly analyze the queueing delay while the transmission • The average rate of loss packets due to the buffer size
delay is treated as a constant and is utilized to determine constraint in the queue of Un can be given by
the value of tw . Finally, the channel can be blocked when
the packets are transmitting. The blockage occurs accord- ∑ ∑n
Cn D+C
Lnbuf f er = λn pnc,d . (11)
ing to a Poisson process with parameter α. Whenever the
c=0 d=0
channel is blocked, it will be recovered after a period of
time. This time is a random variable that is assumed to be • The expected number of the blocked channels in the
distributed according to the negative exponential distribution queueing system of Un can be given by
{ }
with parameter β. Let Pn = Pn0 , Pn1 , . . . , Pnc , . . . , PnCn ∑ ∑n
Cn D+C
denote
{ the steady-state probability}matrix of Un , where Pnc = En [break] = cpnc,d . (12)
pnc,0 , pnc,1 , . . . , pnc,d , . . . , pnc,D+Cn represents the steady-state c=0 d=0

probability vector when there are c blocked servers, and pnc,d • The expected number of channels where no URLLC
represents the steady-state probability when there are d waiting packet is transmitting can be given by
packets as well as c blocked servers in the queueing system. n −1 C
C∑ n −c

By leveraging the matrix geometric method introduced in [37, En [empty] = [(Cn − c) − d] pnc,d . (13)
Chapter 3], we can easily obtain Pn as indicated by the c=0 d=0
following lemma.
Lemma 1: PnCn is the solution to the following equations: In the above, λn is the packet arrival rate of Un . Since all
 Cn channels are assumed to be the same, the average packet
{ n [ n ]}−1

 PCn [ΓCn −(Cn αβ ΓCn −1)− ]Γ0,Cn −2
n n
= 0, dropping probability θn , the blockage probability ψn , and the
C∑n −1 ∏ c (5) idle probability φn of each channel can be derived as θn =

 PCn
n
ϕnj+1 + I e = 1, Lntime + Lnbuf f er En [break] En [empty]
c=0 j=1 , ψn = , and φn = ,
λn Cn Cn
respectively.
where
 n −1
 Γ0,0 = Cn αβ(Γn0 ) , [ n ]−1 D. Problem Formulation
 Γ n
0,c−1 = (C n − c + 1) α (cβ) Γc−1 − Γ n
0,c−2 , (6)
2 ≤ c ≤ Cn − 1, Based on the above analysis, we can obtain the throughput
{ from Em to Bk , denoted by rmk e
. It should be noted that, Un
−1
β(Γn0 ) , j = 0, does not persistently cause interference to Em even if it pairs
n
ϕj+1 = [ ]−1
(j + 1) β Γnj − Γn0,j−1 , 1 ≤ j ≤ Cn − 1. with Em . This is because the spectrum resource of Em is not
(7) reused by Un when there is no URLLC packet transmitting
e
and e represents a column vector whose components all on the channel. Therefore, rmk should consist of the following

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 6

two parts, as be formulated as


M |A
∑ ∑m|
e
rmk
{( = b× ) P1 : max e
rmk ,
∑N ∑ S ∑
N ∑ S {pemk }M ×|Am | ,{puns }N ×S ,{ρmk
ns }M ×|Am |×N ×S m=1
1− ρns ψn −
mk mk e
ρns φn log2 (1 + ξmk )+ k=1
[
n=1 s=1 ( n=1 s=1 )]} (16)
∑N ∑ S pemk gmk,km
e c
gmk b
gkm,mk subject to
ρmk
ns φn log2 1 + . ns ∈ {0, 1} , ∀m, ∀k, ∀n, ∀s,
C1 : ρmk
n=1 s=1 bN0
M |A
∑ ∑ m|

ns = 1, ∀n, ∀s,
(14) C2 : ρmk
e
The first part of rmk represents the throughput when there m=1 k=1
are URLLC packets transmitting on the channel while the ∑N ∑S
e
C3 : ns = 1, ∀k, ∀m,
ρmk
second part of rmk represents the throughput when there is n=1 s=1
only eMBB data transmitting on the channel. ∑S ∑ M |A∑m|

ns , ∀n,
C4 : Cn = ρmk (17)
s=1 m=1 k=1
On the other hand, to support low-latency communications, C5 : 0 ≤ pemk ≤ pe , ∀m, ∀k,
the packet size should be extremely small. Therefore, Shan- C6 : 0(≤ puns ≤ pu), ∀n, ∀s,
non’s capacity formula, which derives the tight upper bound 1
C7 : 1 ns + θn ≤ θ0 , ∀n, s, m, k,
̸= 0 εmk
on the rate at which packets with infinity blocklength can εmk
ns
be reliably transmitted over a communication channel, is no e
C8 : ξmk ≥ ξmin
e
, ∀m, ∀k,
longer appropriate for characterizing the transmission of short-
where pe and pu represent the maximum transmit powers of
size packets. According to [38], the reliability and latency will e
URLLC users and eMBB users, ξmin represents the minimum
be underestimated if Shannon’s capacity is used to analyze
SINR requirement of eMBB users, C2 and C3 guarantee that
the performance of URLLC. This is because we cannot
the URLLC-eMBB user pairing is a one-to-one pairing, Cn in
transmit packets with finite blocklength under arbitrarily small
C4 stands for the number of accessed channels, C5 and C6
error probability. The decoding error probability, ε, has been
represent the power constraints, C7 represents the reliability
approximated in [39] as
constraints, and C8 represents the SINR constraints. Note that
ε = Q (f (γ, Λ)) , (15) the latency of each packet has been strictly constrained in
our queueing model, which is reflected in the average rate of
1 ∫ ∞ − t2
where Q (x) = √ e 2 dt, f (γ, Λ) = loss packets due to the time constraint. Therefore, the time
√ 2π x constraint has already been considered in C7.
Λ −2
ln 2 log2 (1 + γ), and V = 1 − (1 + γ) . In the Problem P 1 is a mixed-integer non-linear programming
V (MINLP) problem, which is very hard to solve in general.
above, Λ is the channel blocklength and γ is the SINR
of the URLLC channel. In this paper, we consider the Meanwhile, one important queueing metric, i.e., the number
effect of short blocklength to characterize the relationship of simultaneous transmission links Cn , is embedded in the
among achievable rate, decoding error probability, and objective function and constraints. However, since interference
transmission delay of URLLC more accurately. Note that exists only within the reusing link of each spectrum reusing
this effect has also been considered in many studies on pair, we can at first investigate the power allocation for
resource management in URLLC systems, such as [7], each possible combination of {n, s, m, k} to maximize the
capacity on eMBB user’s channel while guaranteeing the
ns ξns , ∀n, m, k, s and
mk
[8], [15]. Here,
( ( mk we define
)) ξns = ρmk u

εns = Q f ξns , Λ , ∀n, m, k, s. If ρns = 1, ξns


mk mk mk
and SINR constraint of eMBB user and the reliability constraint
mk
εns equal to the corresponding SINR and the decoding of URLLC user. After that, the resource matching within
error probability, respectively. Otherwise, ξns mk
eqauls to 0 each possible eMBB-URLLC pair, i.e., the determination of
mk
and εns accordingly approaches to infinity. Based on this, the URLLC user on reusing which frequency band of the
we( can derive eMBB user and accessing which mmWave BS, is independent
) the reliability constraint for each channel as of each other. Therefore, the resource matching with each
1
1 mk ̸= 0 εmk ns + θn ≤ θ0 , where θ0 is the maximum possible eMBB-URLLC reusing pair can be solved once we
εns
mk obtain the optimal power allocation. It should be noted that,
( probability. When ρns = 1, the ))
allowed(error inequation turns
u u c b
pns gns,sn gns gsn,ns since Cn is embedded in both power allocation and resource
into Q f ,Λ + θn ≤ θ0 , matching problems, we need to solve the power allocation and
bN0 + pemk gmk,sn
e c gb
gms sn,km resource matching problems under a fixed Cn and then traverse
which means that the summation of the decoding error
all possible Cn to achieve the optimal resource matching.
probability and the packet loss probability cannot exceed
After we obtain the optimal resource matching between each
the maximum allowed error probability. When ρmk ns = 0, the possible eMBB-URLLC pair, the optimal user paring problem
inequation turns to θn ≤ θ0 .
can be effectively solved using the Hungarian method [40].
In this paper, we aim at maximizing the throughput of IV. R ESOURCE M ANAGEMENT A LGORITHM
eMBB users while guaranteeing the latency and reliability In this section, we will propose an efficient algorithm to
requirements of URLLC users. Therefore, the resource man- solve the resource management problem P 1 formulated in
agement problem for the proposed coexistence scenario can Section III. We will first present the solutions to the three

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 7

subproblems, i.e., power allocation, resource matching, and Note that f2 , f2−1 , f3 , f3−1 are all monotonically one-to-
user pairing, respectively. Then we will develop the whole one functions, for which we can find the optimal solution by
algorithm and analyze the computational complexity. standard bisection searching.

A. Power Allocation B. Resource Matching


Assuming ρmk ns = 1, the power allocation problem can be We consider the resource matching for all possible eMBB-
expressed as URLLC pairs in this subsection. Without loss of generality,

P 2 : emaxu f (pemk , puns ) = we focus on the resource matching problem between Un♯
pmk ,pns
( ) and Em♯ . Note that the resource matching is based on the
pemk gmk,km
e c
gmk b
gkm,mk
log2 1 + u c gb
, optimal power allocation, which we have analyzed in Section
bN0 + puns gns,km gnk km,ns IV-A. Since power allocation is different when Cn changes, we
(18) still need to fix Cn to obtain the corresponding solution, and
subject to C5, C6, and then traverse all Cn ∈ {1, 2, ..., |Am |} to achieve the optimal
pemk gmk,km
e c
gmk b
gkm,mk
C9 : u c gb
≥ ξmin
e
, resource matching.
bN0 + puns gns,km gnk km,ns
(19) Denote rne∗♯ ,s,m♯ ,k as the optimal throughput of the trans-
C10 : εmk
ns + θn ≤ θ 0 . mission from Em♯ to Bk , whose operating channel is reused
Note that only C10 is related to Cn . Moreover, the power by Un♯ associated with Bs . Then, rne∗♯ ,s,m♯ ,k can be expressed
allocation problem is independent of the resource matching as
rne∗♯ ,s,m♯ ,k = b×
and the user pairing when Cn is fixed. On the other hand, we ( )
also need power allocation under different Cn as a prerequisite pe∗ e c b
m♯ k gm♯ k,km♯ gm♯ k gkm♯ ,m♯ k
[(1 − ψn♯ − φn♯ ) log2 1 +
of the resource matching problem. Therefore, we will solve the bN0 + pu∗ g u s,km gnc ♯ k gkm,n
b
( n♯ s n♯)] ♯s

power allocation for all possible Cn ∈ {1, 2, ..., |Am |}. pe gme c b
♯ k,km♯ gm♯ k gkm♯ ,m♯ k
When Cn is fixed, θn becomes a constant. Therefore, the + φn♯ log2 1 + ,
bN0
constraint C10 turns into
ns ≤ θn
C11 : εmk = θ0 − θn∗ .
Cn ∗ ∆
(20) ( ) (23)
where term pu∗ ♯ g u
♯ g c

n s n s,km n k km,n s g b
♯ represents the interfer-
We then define function f2 : R → R ence from the transmission of URLLC user Un♯ to eMBB user
f2 (puns ) gmk,km
e c
gmk b
gkm,mk e∗
satisfying = e
ξmin , Em♯ . On the other hand, denote rm ♯ ,k as the throughput from
u
bN0 + puns gns,km c gb
gnk km,ns Em♯ to Bk without the reusing of URLLC users, which can
and( ( function f3 : R → )) R satisfying be expressed as
u
puns gns,sn c b
gns gsn,ns ∗ rme∗
♯ ,k =
Q f e c gb
,Λ = θnCn . ( )
bN0 + f3 (puns ) gmk,sn gms sn,km pe gme c b
♯ k,km♯ gm♯ k gkm♯ ,m♯ k

The following corollary shows whether there exists a feasible b (1 − ψm♯ ) log2 1 + .
bN0
solution to the power allocation problem. (24)
Corollary 1: The power allocation problem has feasible Then, the throughput loss due to channel reusing can be
 ∗ conditions are satisfied:
solutions when the following achieved as
 θn ≤ θe0
♯ ,k − rn♯ ,s,m♯ ,k .
e∗ e∗
Gn♯ ,s,m♯ ,k = rm (25)
0 ≤ p mk ≤ pe ,

(21)
 ⌢u
0 ≤ p ns ≤ pu Denote xk,s = ρm
∆k ♯ ∆
n♯ s and yk,s = Gn♯ ,s,m♯ ,k , the resource
( e ) ( u ) ( u ) matching problem for Un♯ and Em♯ can be expressed as
⌢ ⌢u ⌢ ⌢
where p mk , p ns satisfies f2 p ns = f3 p ns =
( ) |A
∑ m♯ | S

⌢e ⌢e ⌢u ∆
p . Note that if there is no feasible p mk , p ns , we set P 3 : fn♯ ,m♯ = min xk,s yk,s , (26)
( mk ) xk,s
⌢e ⌢u k=1 s=1
p mk , p ns to (+∞, +∞).
subject to |A∑
m♯ |
The proof of the corollary can be easily obtained
by justifying whether the intersection of f2 and f3 C12 : xk,s ≤ 1, ∀s,
is located in the feasible power allocation area P = k=1

S
{(pemk , pumk ) |0 < pemk < pe , 0 < puns < pu }. C13 : xk,s ≤ 1, ∀k,
s=1 (27)
The following lemma characterizes the optimal solution to
P 2, whose proof can be found in Appendix A. C14 : xk,s = {0, 1} , ∀s, ∀k,
∑S |A∑m♯ |
Lemma 2: The optimal power allocation can be expressed
C15 : xk,s = Cn .
as  ( e −1 e ) s=1 k=1

 p {[(
, f3 (p ) , )] }


 [( f2 −1 (pu ) < pu) or ]
Note that if the power allocation problem of ρmkns = 1 has no
if ,
(pe∗ u∗
mk , pns ) = f2−1 (pe ) > pu and (f3 (pu ) > pe ) feasible solution, we set yk,s as a sufficiently large positive

 (f3 (pu ) , pu ) ,

 number to avoid the condition that xk,s = 1.
 [( −1 u ) ]
We can see that P 3 is a combinatorial problem. Unfortu-
if f2 (p ) > pu and (f3 (pu ) < pe ) .
(22) nately, it cannot be solved through the widely-used Hungarian

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 8

algorithm due to the constraint C15. In the following, we will îï


ì
m

first transform it into a minimum cost network flow problem


and then develop an algorithm to obtain the optimal solution.
ï
ï
ï
Moreover, we also present a low-complexity algorithm, which í
is based on the matching theory. w

í
ï
S ï
a) The optimal solution: Minimum cost network flow:
ï
ì
Before we proceed, we first briefly review some preliminaries
of the minimum cost network flow problem. The problem
consists of finding the minimum cost way of sending a certain ïî
amount of flow from the supply node to the demand node ì
through a flow network. Some parameters of the flow network ï
ï
are known, including the demand for the flow of some nodes, = S - Cn í
'
m
z
the lower and upper bounds on the amount of the flow that ï
each edge in the flow network can undertake, and the cost for ïî
transporting one unit of the flow through every edge. We aim
to find the optimal flow allocation to determine how many Fig. 3: The graph of the resource matching problem.
units of the flow that each edge will undertake to minimize
the sum cost while meeting the constraints. as ∑
Now we will show how to transform P 3 into a minimum P 5 : min t (j) x (j), (30)
′ x
cost network flow problem. First, we build a set Am♯ consist- j∈J
ing of S − Cn virtual eMBB communication links. Based on subject to
this, the original problem can be converted into
C17 : 0 ≤ x (j) ≤ 1, ∀j ∈ J ,
∑ ∑
S ∑ ′

P 4 : min xk,s yk,s , (28) C18 : x (jn1 ,n3 ) + x (jn3 ,z ) = 0, ∀n3 ∈ Am♯ ,
xk,s n1∑∈V1

k∈Am♯ ∪A
m♯
s=1 C19 : x (jn1 ,n2 ) = 1, ∀n1 ∈ V1 ,
n1 ∈V
∑1 (31)
subject to C13, C14, C15, and C20 : x (jn2 ,w ) = Cn ,
∑ n2 ∈Am♯

C16 : xk,s = 1, ∀s, (29) ∑

C21 : x (jn3 ,z ) = Am♯ − Cn .
k∈Am♯ ∪A ′
m♯ n3 ∈A
m♯

where yk,s is set to be 0 if k ∈ Am♯ . Then, we construct To solve P 5, we have the following theorem, whose proof
a directed graph G = (V, J ) as shown in Fig. 3, where can be found in [41].
V is the node set and J is the edge set. The set of nodes Theorem 1: The minimum cost network flow problem al-
can be divided as V = V1 ∪ V2 ∪ V3 , where V1 = S, ways has an optimal integral solution if

V2 = Am♯′ ∪ Am♯′ , and V3 = {w, z}. Compared with the • There exists an optimal solution.
bipartite graph representing the assignment problem without • The upper bounds and lower bounds of edges are integral.
subset constraints, the third layer V3 is augmented to represent • The demands of nodes are all integral.
subset constraint C15. Specifically, we set w to represent the
set of real communication links, and z to represent the set of Generally, the minimum cost network flow problem can
virtual communication links. Therefore, we add an edge from be solved with polynomial computational complexity. The
′ most famous algorithm is the Ford-Fulkerson algorithm [42].
each node in Am♯ to w, and from each node in Am♯ to z. We
utilize jn1 ,n2 to represent the edge from node n1 to node n2 . However, sometimes the convergence cannot be guaran-
Then, we assign demand −1 to each node in V1 , demand teed. The Edmonds-Karp algorithm [43], revising the Ford-
′ 0 to
′ (
Fulkerson )
algorithm, is with the computational complexity
each node in Am♯ , demand Cn to w, and demand Am♯ − Cn 2
of O |V| |J | . Its computational complexity can also be
to z. The total number of nodes in G is |V| = 2K + 2. Finally, ( )
5 2
we assign the cost t(j) of the edge jn1 ,n2 as yn1 ,n2 , and the expressed as O |V| since |J | ∝ |V| , which is too high for
cost of edges jn1 ,n3 , jn2 ,w , and jn3 ,z as 0, where n1 ∈ V1 , practical implementation. Therefore, we adopt the cost scaling

n2 ∈ Am♯ , and n3 ∈ Am♯ . Denote x(j) as the allocated flow based push-relabel algorithm to (solve)P 5, whose computa-
3
for edge j. According to the constraint C14, we can set the tional complexity is reduced to O |V| . The algorithm main-
upper bound and low bound of x(j) as 1 and 0, respectively. tains a “preflow” and gradually converts it into a maximum
Now we have the following three essential constraints: the flow by moving flow locally between neighboring nodes using
demand for the flow of some nodes, the lower and upper push operations under the guidance of an admissible network
bounds on the amount of the flow that each edge in the flow maintained by relabel operations. The detailed procedure of
network can undertake, and the cost for transporting one unit the push-relabel algorithm can be found in [44].
of the flow through every edge. Based on this, problem P 4 can b) A low-complexity algorithm: Student project alloca-
be transformed into a minimum cost network flow problem P 5 tion: Thus far, we have developed an algorithm with cubic

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 9

computational complexity to find the optimal solution of P 4. the objective function in P 6 is related to the choices of
However, considering that we need to iterate |Am | times for both mmWave BSs and channels, it is appropriate to con-
1 struct a SPA model with preferences over the (BS, channel)
different Cn , and there are in total N M possible URLLC-
2 pair. Therefore, we adopt the SPA with preferences over
eMBB pairs in the system, we may consume a high computa-
the (student, project) pair (SPA-(S,P)) model. In our model,
tional time to find the optimal resource matching solutions for
the preferences of L1 (Lecture 1) and BSs (students) are
all possible pairs. Therefore, it is very important to develop
constructed as the positive order of yk,s , and the preference
an algorithm with low computational complexity.
of L2 (Lecture 2) is constructed as the reverse order of yk,s .
The proposed low-complexity algorithm is based on the This is because we want to allocate high throughput to realistic
matching theory. Different from traditional optimization meth- channels and low throughput to virtual channels.
ods that aim to achieve the optimality, the matching the- Denote the set of the BSs that Un is associated
ory mainly considers stability. Specific to resource matching with as BS n = (BS1n , . . . , BSsn , . . . , BSSn ) and the set
problem P 4, the stability implies that swapping any two (BS
) from Em as CH
m
(
of the possible reused channels =
association, reused channel) pairs cannot benefit both simul- m m m
CH1 , . . . , CHk , . . . , CH|Am | . The following definition
taneously. From the perspective of optimization, matching
theory simplifies the combinatorial optimization problems by of the block pair is necessary for defining the stability of the
transforming the relations among BS associations and reused matching.
Definition 1: The pair (BSsn , CHkm ) ∈ BS × CH, where
channels from values to preference rankings. Although stabil-
CHkm ∈ Li , blocks a matching M if:
ity is only the necessary condition of the optimality, matching
• BSs prefers CHk to M (BSs ).
n m n
theory-based algorithms are time-efficient and can achieve a
• Either
performance close to the optimal solution. In addition, the
matching theory has the ability to inherently capture suitable – CHkm is free and either
models for various wireless communication systems. In this ∗ M (BSsn ) ∈ Li and li prefers (BSsn , CHkm ) to
paper, we adopt the student project allocation (SPA) model (BSsn , M (BSsn )), or
[45], which is one of the most famous matching theory models, ∗ M (BSsn ) ∈ / Li and li is under-subscribed, or
to solve P 4. In the following, we first relax the original ∗ M (BSsn ) ∈ / Li , li is full, and li prefers
problem, present the SPA model, and then introduce the SPA- (BSsn , CHkm ) to the worst pair (BSsn′ , CHkm′ )
based algorithm. Finally, we will show that the relaxation will that is being assigned to lk , or
not affect the stable solutions to the original problem P 4. – CHkm is matched and li prefers (BSsn , CHkm ) to
We define the relaxed resource matching problem P 4 as the pair (BSsn′ , CHkm ), where BSsn′ is the worst BS
association choice in M (CHkm ) and either
∑ ∑
S
P 6 : min xk,s yk,s , (32) ∗ M (BSsn ) ∈ / Li , or
xk,s
′ s=1 ∗ M (BSsn ) ∈ Li and and li prefers (BSsn , CHkm )
k∈Am♯ ∪A
m♯ to the pair (BSsn , M (BSsn )).
subject to C13, C14, C16, and Definition 2: M is stable if and only if there is no blocking
pair of it.
S |∑
∑ m♯ |
A
Based on the SPA-(S,P) formulation, the preference con-
C22 : xs,k ≤ Cn . (33)
struction, and the definition of the stability, we can apply
s=1 k=1
the SPA-(S,P) algorithm, whose detail is presented in [45], to
Note that the only difference between P 4 and P 6 is that C15 obtain the stable solutions of P 6. The basic rule of the SPA-
is changed into C22. (S,P) algorithm is to follow the principle of the classical Gale-
The SPA model is a variation of the fundamental two- Shapley algorithm [46]. The ( computational
) complexity of the
sided matching model, where various students are assigned to SPA-(S,P) algorithm is O S 2 . Then we have the following
various projects owned by different lectures. Compared with lemma to show the relation between P 4 and P 6, which can
the two-sided matching model, the most important feature of be proved through reduction to absurdity shown in Appendix
the SPA model is that the projects are divided into several sets B.
according to the lecture they are associated with. Moreover, Lemma 3: The stable solutions to P 4 are also to P 6.
each lecture has its own capacity constraint. Therefore, we In other words, the relaxation in C22 does not impact the
regard the potential mmWave BSs that the URLLC user solutions from the perspective of the matching theory.
associated with as students, the realistic and virtual channels
of eMBB users as projects. Meanwhile, we divide the projects C. User Pairing
into two subsets as two lectures, i.e., realistic channel set L1 Thus far, we have developed the optimal power allocation
and virtual channel set L2 , which can be regarded as two for each reused uplink channel and the optimal resource
lectures. According to constraint C22, the capacity of L1 is matching for each eMBB-URLLC pair. In this subsection, we
set to be Cn . focus on user pairing among eMBB and URLLC users.
According to different types of preferences, the SPA model ∑S |A
∑ m|

has several variations, e.g., SPA with preferences over projects Denote πn,m = ρmk
ns as the user pairing indicator,
s=1 k=1
(SPA-P), SPA with preferences over students (SPA-S). Since i.e., πn,m = 1 if Un and Em are paired, πn,m = 0 otherwise.

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 10

TABLE II: Simulation Parameters.


Based on the optimal power allocation and resource matching,
Parameters Typical values
the user pairing problem can be rewritten as
Radius of mmWave SBS 100 m

N ∑
M
∗ Carrier frequency/bandwidth 28 GHz/2 GHz
P 7 : max πn,m fn,m , (34)
πn,m Total mmWave SBS Tx power 37 dBm
n=1 m=1
Path loss model 157.4 + 32 log (R), R in km
subject to Noise spectral density −134 dBm/Hz

M
C23 : ρn,m = 1, ∀n, Shadowing 12 dB
m=1
(35) Operating beamwidth of mmWave SBS 5◦
∑N
10◦
C24 : ρn,m ≤ 1, ∀m. Operating beamwidth of user
n=1 Side lobe gain 0.1

Here we set fn,m = −∞ if there is no feasible resource
matching solution between URLLC user n and eMBB user m. 5.5

Average rate of eMBB users (Mbps/Hz/user)


By adding virtual URLLC users, problem P 7 can be efficiently
solved by the Hungarian method in polynomial time. 5.0

D. Algorithm Design and Computational Complexity Analysis


4.5
In summary, the detailed procedures of the proposed algo-
rithm are summarized in Fig. 4. The computational
( ) complexity
of each power allocation procedure is O D3 . The com- 4.0

(
putational )
complexity of each resource matching procedure
( ) The scenario with no reusing
3
is O (2S) for the optimal solution and O S 2 for the 3.5 The sub-optimal algorithm
sub-optimal solution. The computational complexity of the The optimal algorithm
( )
user pairing procedure is O M 3 . Therefore, the compu-
3.0
tational complexity of the( optimal ( 3algorithm )and the ) sub- 0 2 4 6 8 10
2 2 3 3
optimal
( algorithm
( ) are O M ) N D S + 8S + M and The number of URLLC users
O M 2 N D3 + 1 S 2 + M 3 , respectively.
Fig. 5: Average rates of eMBB users with different number of
V. S IMULATION R ESULTS URLLC users.
In this section, we evaluate the performance of the proposed
algorithms. We consider a mmWave communication network reliability constraint of Un . Meanwhile, we leverage the user
ranged in a square area of 100 m × 100 m, where all mmWave association algorithm developed in our previous work [30] to
BSs, URLLC users, and eMBB users are uniformly distributed. find the appropriate channel selection for URLLC users. In the
The multi-connectivity functionality is enabled for both the simulation, we set N = 10, M = 20, tw = 0.1 ms, θ0 = 10−5 .
URLLC and eMBB users. Each eMBB user has already Furthermore, we set pu = 23 dBm for all URLLC users and
been associated with several mmWave BSs while the optimal pe = 23 dBm for all eMBB users. We also assume that each
association strategy of eMBB users for maximizing the system eMBB and URLLC user can access at most 4 mmWave BSs
throughput can be obtained by the algorithms proposed in simultaneously.
[30]. For simplicity, it is assumed that the directions of the The test results are shown in Fig. 5. From the figure, the
transmission beams between users and mmWave BSs coincide proposed coexistence scenario significantly outperforms the
with the mutual geographical boresight connections. For the orthogonal scenario, especially when the number of URLLC
queueing system, the service time of each packet of all users is large. It is because that the reuse strategy in our
mmWave channels are exponentially distributed with mean proposed system can make eMBB and URLLC users share the
t1 = 10−2 ms, the blockage of each mmWave channel occurs spectrum while guaranteeing the requirements of both, and the
with a Poisson breakdown rate α = 0.1, and the recovery spectrum resource can be utilized more efficiently. Meanwhile,
time of the blocked channel is distributed according to the we can see that the performance of the sub-optimal algorithm
exponential distribution with parameter t2 = 10−2 ms. The is very close to that of the optimal algorithm, validating the
simulation parameters are shown in Table II unless otherwise effectiveness of the proposed suboptimal algorithm.
stated. Then, we show the impact of the multi-connectivity func-
First, we aim to validate the effectiveness of our proposed tionality on the proposed coexistence scenario. Fig. 6 depicts
coexistence scenario. As a comparison, we also test the the number of URLLC users that can be served under the
performance of the orthogonal scenario, where URLLC users scenarios with different λn , tw , θ0 , and different numbers of
with arrival packets select several channels of eMBB users maximum permissible multi-connectivity links. In the simula-
and then exclusively occupy it until there is no packet in its tion, we set M = 20 and N = 10. Further, we set pu = 23
buffer. To maximize the throughput of eMBB users in this dBm for all URLLC users and pe = 23 dBm for all eMBB
scenario, URLLC user Un will utilize an channels, where users. To better show the results, we assume that both eMBB
an is the minimum number of channels to guarantee the and URLLC users have the same numbers of maximum multi-

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 11

 
    ! "#$%&'( U1 )*+ E1

C1 = 1 C1 = s = 1: S
s = 1: S 1

k = 1: k = 1:
,-./0 123456789: ;<= 1
XYZ[\ ]^_`abcd`e fgh ijklm nopqrstuvw xyz 1
{|}~ €‚ƒ„…†‡ƒˆ ‰Š‹
s = 1, k = 1 s = S,k = 1 C1 = 1: 1 s = 1, k = 1 s = S,k = 1

>?@ABCDE FGHIJKLM ŒŽ‘’“ ”•–—˜™š›


NOPQRST U1 UVW E1 œžŸ ¡¢ U1 £¤¥ E1

î
ï
¦§¨©ª«¬ ­®¯°±²³´µ ¶· ¸¹º C1 »¼ ½¾¿ÀÁ ÃÄÅÆÇÈÉ ð
ÊËÌÍÎ ÏÐÑÒÓÔÕÖ×Ø ÙÚÛ ÜÝÞßàáâã äåæçèéêë ñ
ìíîïðñò U1 óôõ E1 ò
ó
ô
õö
m = 1: M , n = 1: N ÷
øù
ö÷øùú ûüýþÿ   
   U n  ! Em
ú
û
Cn = 1 s = 1: S Cn = ü
m s = 1: S ý
k = 1: m k = 1: m þ
"#$%& '()*+,-./0 123 NOPQR STUVWXYZ[\ ]^_ `abcd efghijklmn opq rstuv wxyz{|}~€ ‚ƒ
s = 1, k = 1 s = S,k = m Cn = 1: s = S,k = m ÿ
m s = 1, k = 1


456789:; <=>?@ABC „…†‡ˆ‰Š‹ ŒŽ‘’“




DEFGHIJ U n KLM Em ”•–—˜™š U n ›œ Em




žŸ ¡¢£¤ ¥¦§¨©ª«¬­ ®¯ °±² Cn ³´ µ¶·¸¹º »¼½¾¿ÀÁ


ÂÃÄÅÆ ÇÈÉÊËÌÍÎÏÐ ÑÒÓ ÔÕÖ×ØÙÚÛ ÜÝÞßàáâã
äåæçèéê U n ëìí Em

Fig. 4: The flowchart of the proposed algorithm.

connectivity links, and each URLLC has the same packet in the queueing system. Secondly, a smaller tw would lead to
arrival rate in the simulation. From the figure, only less than more stacked packets in the buffer, thus increases the total
80 percent of users can be served in all simulated scenarios packet drop probability, which can also be worked out by
even if λn = 1000 packets/s. This is because that mmWave enabling more URLLC transmission links.
channels can be blocked when encountering solid obstacles,
including the blockage rate of each mmWave channel and Fig. 7 shows the impact of the error probability, the maxi-
the recovery time of the blocked channel. Without multi- mum power, and the latency requirements on the performance
connectivity, there is no other available channel so that the of eMBB users. In the simulation, we set M = 20, N = 10,
packets cannot be transmitted when the serving channel is λn = 4000 packets/s, and pu = 23 dBm for all URLLC
blocked. As a result, more packets will be stacked in the users. Meanwhile, we also set the maximum multi-connectivity
buffer, thus will be dropped more often. To deal with this issue, links as 4. From the figure, the performance of eMBB users
more channels should be allocated for URLLC users. This is increases as the error probability decreases. The reasons are
the reason why multi-connectivity should be introduced into two-fold. First, according to (15), the decoding error prob-
mmWave URLLC. From the figure, all simulated scenarios ability of a URLLC user decreases as the SINR increases.
can handle almost all URLLC users with λn = 1000 packets/s Therefore, to guarantee a more stringent error probability
when two multi-connectivity links are enabled. For users with requirement, a URLLC user may increase its transmit power
heavier packet load, we must enable more multi-connectivity to obtain a higher SINR, which will degrade the performance
links to support the URLLC service. This result shows that of eMBB users. Secondly, as we have stated before, a URLLC
five multi-connectivity links can serve all URLLC users with a user can enable multi-connectivity functionality, i.e., reusing
packet arrival rate under λn = 4000 packets/s in all simulated more channels from eMBB users, to reduce the packet drop
scenarios while the ratio will drop to 90% when enabling probability. This also has a negative impact on the performance
four multi-connectivity links. Besides, we can also see that of eMBB users. In addition, one can observe an interesting
the simulated scenarios with more stringent constraints on tw result that the average rate of eMBB users and the logarithm of
and θ0 need more multi-connectivity links to serve URLLC the error probability have a nearly linear relationship. Besides
users. The reasons can be elaborated in the following. First, a the error probability, we can also see that the average rate of
lower θ0 implies a lower per-channel packet drop probability, eMBB users decreases as the latency requirement increases. It
whereas increasing the number of multi-connectivity links can is because that, a more tight latency requirement will lead
reduce the total packet drop probability by dint of more servers to a higher packet drop probability under the same SINR.
As we have discussed earlier, two approaches are effective to

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 12

10 10

9 9

The number of URLLC users


The number of URLLC users

8 8

7 7

6 6
=1000 n
=1000
n

=2000 5 =2000
5 n n

=3000 n
=3000
n
4 4
=4000 n
=4000
n

3 3
1 2 3 4 5 6 1 2 3 4 5 6

The number of maximum multi-connectivity links The number of maximum multi-connectivity links

(a) tw = 0.1 ms, θ0 = 10−5 . (b) tw = 0.05 ms, θ0 = 10−5 .

10 10

9 9
The number of URLLC users

8 The number of URLLC users 8

7 7

6 6
n
=1000 n
=1000
5 n
=2000 5 n
=2000
n
=3000 n
=3000
4 4
n
=4000 n
=4000
3 3
1 2 3 4 5 6 1 2 3 4 5 6
The number of maximum multi-connectivity links The number of maximum multi-connectivity links
(c) tw = 0.1 ms, θ0 = 10−6 . (d) tw = 0.05 ms, θ0 = 10−6 .
Fig. 6: The number of accessible URLLC users with different packet arrival rates and different numbers of maximum multi-
connectivity links.

maintain the error probability: increasing the power to obtain


6.0
higher SINR for lower decoding error probability or reusing
Average Rate of eMBB users (Mbps/Hz/user)

more channels from eMBB users to decrease the packet


dropping probability. Both can decrease the performance of
5.5
eMBB users. Finally, the average rate of eMBB users increases
with pe , which is rather straightforward. However, we can see
that in the scenario with pe = 25 dBm and tw = 0.1 ms,
the degradation caused by stringing the latency limit is much 5.0
bigger than that caused by reducing the maximum achievable
power. This phenomenon demonstrates the importance of the pe=23 dBm, tw=0.1 ms
latency requirement in the URLLC system. 4.5 pe=25 dBm, tw=0.1 ms
pe=23 dBm, tw=0.05 ms
VI. C ONCLUSION pe=25 dBm, tw=0.05 ms
This paper develops a novel URLLC-eMBB nonorthogonal 4.0
coexistence scenario, where URLLC users reuse the channels 1E-7 1E-6 1E-5 1E-4 1E-3 0.01 0.1 1

of eMBB users to access multiple mmWave BSs. We fully The error probability ε
utilize the unique characteristics of mmWave communications,
Fig. 7: Average rates of eMBB users with different power,
i.e., beamforming and multi-connectivity functionality, to en-
latency and reliability constraints.
hance the performance and robustness of the coexistence sce-

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 13

nario. A resource management problem is formulated to max- (ηpemk , ηpuns ) satisfies both C9 and C11, which is proved as
imize the throughput of eMBB users while guaranteeing the a feasible solution pair.
delay and reliability requirements of URLLC users. Through Now we turn to the objective function f (pemk , puns ). As in
appropriate problem decomposition, an optimal algorithm is (36), we can prove that f (ηpemk , ηpuns ) is a monotonically
developed. Besides, we also present a low computational com- increasing function of η, which means that there is at least
plexity algorithm, which achieves a near-optimal performance. one power in the optimal solution (pe∗ u∗
mk , pns ) that can achieve
Simulation results show that our proposed coexistence scenario its upper bound.
can achieve much better performance than the scenario without The feasible set of the power allocation problem can be
reusing. Meanwhile, the necessity of multi-connectivity for illustrated by the two-dimensional plane in Fig. 8, where the
guaranteeing the reliability of URLLC is validated. Our results horizontal and vertical axes stand for puns and pemk , respec-
in this paper serve as a valuable reference for the design of tively. The points beyond solid line f2 can guarantee C9, while
future URLLC systems. the points under solid line f3 can guarantee C11. Therefore,
Although we have presented the system design and resource the areas filled with slashes represent the feasible areas of
management of the mmWave URLLC system, there are still the power allocation problem. With different relationships
several interesting problems in this area. Our future works may between pu and pe , there are totally three cases. Since there is
consider a more complicated scenario, where each URLLC at least one power in the optimal solution (pe∗ u∗
mk , pns ) achieves
user can reuse the resource from multiple eMBB users while its upper bound, we can easily find that the black points sitting
each eMBB user can share its resource with multiple URLLC on the edges of the feasible areas are the optimal points for
users. It is very interesting to explore whether this complicated all cases.
reusing strategy can further improve the data rate of eMBB A PPENDIX B
users and the reliability of URLLC users. However, resource P ROOF OF L EMMA 3
management in this scenario is also very challenging. Besides,
If Lemma 3 does not hold, there will be at least one stable
it is significant to study the coexistence scenarios where the
matching M̃ violating C15. Since Cn ≤ |Am♯ | ≤ S, there
packet retransmission, fast fading channels, or partial/non-
must exists at least one realistic channel that is unmatched,
channel state information is considered. Moreover, a more gen-
and at least one BS that is matched with a virtual channel.
eral coexistence scheme among eMBB, URLLC, and MTC, is m
also an interesting open issue. n m
) is CHk , andnthe (BS,
Suppose the unmatched(realistic channel
virtual channel) pair is BSs , CHk♮ . Obviously, BSs prefers
A PPENDIX A CHkm to M f (BS n ) = CH m♮ , and CH m is free. According
s k ( k )
P ROOF OF L EMMA 2 to Definition 1, we can see that BSsn , CHkm♮ is a block
We first prove that for any given feasible solution pair f which makes M
pair in M, f unstable. This contradicts the
(pemk , puns ), (ηpemk , ηpuns ) is always a feasible solution when assumption that M f is a stable matching. Therefore, Lemma 3
pemk gmk,km
e c
gmk b
gkm,mk always holds.
η > 1. Obviously, if u c gb
≥ ξmin
e
, we
bN0 + puns gns,km gnk km,ns
ηpemk gmk,km
e c
gmk b
gkm,mk R EFERENCES
have u c gb
≥ ξmin
e
since [1] P. Popovski, “Ultra-reliable communication in 5G wireless systems,” in
bN0 + ηpuns gns,km gnk km,ns Proc. 1st Int. Conf. 5G Ubiquitous Connectivity, pp. 146-151, Nov. 2014,
Kittilä, Finland.
ηpemk gmk,km
e c
gmk b
gkm,mk pemk gmk,km
e c
gmk b
gkm,mk [2] Study on Scenarios and Requirements for Next Generation Access Tech-
u c gb
=
bN0 + ηpuns gns,km gnk bN0 u c gb
nologies, V14.2.0, document TR 38.913, 3GPP, Jun. 2017.
km,ns + puns gns,km gnk km,ns [3] H. Ji et. al., “Ultra-reliable and low-latency communications in 5G
η downlink: Physical layer aspects,” IEEE Wireless Commun., vol. 25, no.
pemk gmk,km
e c
gmk b
gkm,mk 3, pp. 124-130, Jun. 2018.
> u c gb
. [4] S. R. Pandey, M. Alsenwi, Y. K. Tun, and C. S. Hong, “A downlink
bN0 + puns gns,km gnk km,ns resource scheduling strategy for URLLC traffic,” in Proc. Int. Conf.
(36) BigComp, Feb. 2019, Kyoto, Japan.
Then we focus on the constraint C11. According to [5] S. Xing, X. Xu, Y. Chen, Y. Wang, and L. Zhang, “Advanced grant-
(15), it is easy to prove that εmk ns is a monotonically free transmission for small packets URLLC services,” in Proc. ICC
Workshops, May 2019, Shanghai, China.
puns gns,sn
u c b
gns gsn,ns [6] D. Wu, J. Wang, Y. Cai, and M. Guizani, “Millimeter-wave multime-
decreasing function of e e c gb
.
bN0 + pmk gmk,sn gms sn,km
dia communications: Challenges, methodology, and applications,” IEEE
With the same operations as (36), we can Commun. Mag., vol. 53, no. 1, pp. 232-238, Jan. 2015.
[7] C. Sun et. al., “Optimizing resource allocation in the short blocklength
ηpuns gns,sn
u c b
gns gsn,ns regime for ultra-reliable and low-latency communications,” IEEE Trans.
easily prove that >
bN0 + ηpemk gmk,sn e c gb
gms sn,km
Wireless Commun., vol. 18, no. 1, pp. 402-415, Jan. 2019.
[8] C. She, C. Yang, and T. Q. S. Quek, “Joint uplink and downlink resource
puns gns,sn
u
gnsc b
gsn,ns configuration for ultra-reliable and low-latency communications,” IEEE
e e c b
, which means Trans. Commun., vol. 66, no. 5, pp. 2266-2280, May 2018.
bN0 + pmk gmk,sn ( gms gsn,km ) [9] H. Yang, K. Zheng, L. Zhao, and L. Hanzo, “Twin-timescale radio
ηpuns gns,sn
u c b
gns gsn,ns resource management for ultra-reliable and low-latency vehicular net-
that εmk
ns e e c gb
< works,” IEEE Trans. Veh. Technol., vol. 69, no. 1, pp. 1023-1036, Jan.
bN0 + ηpmk gmk,sn gms
( ) sn,km 2020.
puns gns,sn
u c b
gns gsn,ns [10] A. Anand and G. de Veciana, “Resource allocation and HARQ opti-
εmk
ns e e c gb
. To sum up, mization for URLLC traffic in 5G wireless networks,” IEEE J. Sel. Areas
bN0 + pmk gmk,sn gms sn,km Commun., vol. 36, no. 11, pp. 2411-2421, Nov. 2018.

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 14

e e e
pmk (pe*
(
, pnsu* ) = p e , f 3-1 ( p e ) ) pmk ( pmke* , pnsu* ) = ( p e , f3-1 ( p e ) ) pmk
mk

e e
(p e*
mk (
, pnsu* ) = f 3 ( p u ) , p u )
e
p p p

f 3 ( pnsu )
f 3 ( pnsu ) f 3 ( pnsu )
  

 

f 2 ( pnsu )
f 2 ( pnsu ) f 2 ( pnsu )

0 pu pnsu 0 pu pnsu 0 pu pnsu


f 2
-1
(p )< p
e u
f 2
-1
(p )> p
e u
and f 3 ( p u
)> p e
f 2
-1
(p )> p
e u
and f 3 ( p u
)< p e

Fig. 8: Optimal power allocation.

[11] S. E. Elayoubi, P. Brown, M. Deghel, and A. Galindo-Serrano, “Radio [27] W. Wang et. al., “Sub-channel allocation for device-to-device underlay-
resource allocation and retransmission schemes for URLLC over 5G ing full-duplex mmWave small cells using coalition formation games,”
networks,” IEEE J. Sel. Areas Commun., vol. 37, no. 4, pp. 896-904, IEEE Trans. Veh. Technol., vol. 68, no. 12, pp. 11915-11927, Dec. 2019.
Apr. 2019. [28] Y. Xu, H. Shokri-Ghadikolaei, and C. Fischione, “Distributed association
[12] Z. Hou, C. She, Y. Li, L. Zhuo, and B. Vucetic, “Prediction and com- and relaying with fairness in millimeter wave networks,” IEEE Trans.
munication co-design for ultra-reliable and low-latency communications,” Wireless Commun., vol. 15, no. 12, pp. 7955-7970, Dec. 2016.
IEEE Trans. Wireless Commun., vol. 19, no. 2, pp. 1196-1209, Feb. 2020. [29] C. Zhao, Y. Cai, A. Liu, M. Zhao, and L. Hanzo, “Mobile edge
[13] C. Xiao et. al., “Downlink MIMO-NOMA for ultra-reliable low-latency computing meets mmWave communications: Joint beamforming and
communications,” IEEE J. Sel. Areas Commun., vol. 37, no. 4, pp. 780- resource allocation for system delay minimization,” IEEE Trans. Wireless
794, Apr. 2019. Commun., vol. 19, no. 4, pp. 2382 - 2396, Apr. 2020.
[14] H. Ren, C. Pan, Y. Dong, M. Elkashlan, and A. Nallanathan, “Joint [30] R. Liu, G. Yu, and G. Y. Li, “User association for ultra-dense mmWave
power and blocklength optimization for URLLC in a factory automation networks with multi-connectivity: A multi-label classification approach,”
scenario,” IEEE Trans. Wireless Commun., vol. 19, no. 3, pp. 1786-1801, IEEE Wireless Commun. Lett., vol. 8, no. 6, pp. 1579-1582, Dec. 2019.
Mar. 2020. [31] R. Liu, M. Lee, G. Yu, and G. Y. Li, “User association for millimeter-
[15] J. Tang, B. Shim, and T. Q. S. Quek, “Service multiplexing and revenue wave networks: A machine learning approach,” IEEE Trans. Commun.,
maximization in sliced C-RAN incorporated with URLLC and multicast vol. 68, no. 7, pp. 4162-4174, Jul. 2020.
eMBB,” IEEE J. Sel. Areas Commun., vol. 37, no. 4, pp. 881-895, Apr. [32] T. K. Vu, M. Bennis, M. Debbah, and M. Latva-sho, “Joint path
2019. selection and rate allocation framework for 5G self-backhualed mm-wave
[16] M. Alsenwi, N. H. Tran, M. Bennis, A. K. Bairagi, and C. S. Hong, networks,” IEEE Trans. Wireless Commun., vol. 18, no. 4, pp. 2431-2445,
“eMBB-URLLC resource slicing: A risk-sensitive approach,” IEEE Com- Apr. 2019.
mun. Lett., vol. 23, no. 4, pp. 740-743, Apr. 2019. [33] T. K. Vu et. al., “Ultra-reliable and low latency communication in
[17] Q. Chen, H. Jiang, and G. Yu, “Service-oriented resource management mmWave-enabled massive MIMO networks,” IEEE Commun. Lett., vol.
in spatial reuse-based C-V2X networks,” IEEE Wireless Commun. Lett., 21, no. 9, pp. 2041-2044, Sep. 2017.
vol. 9, no. 1, pp. 91-94, Jan. 2020. [34] H. Shokri-Ghadikolari, C. Fischione, G. Fodor, P. Popovski, and M.
[18] H. Zhang et. al., “Energy efficient user association and power allocation Zorzi, “Millimeter wave cellular networks: A MAC layer perspective,”
in millimeter-wave-based ultra dense networks with energy harvesting IEEE Trans. Commun., vol. 63, no. 10, pp. 3437-3458, Oct. 2015.
base stations,” IEEE J. Sel. Areas Commun., vol. 35, no. 9, pp. 1939- [35] D. Han, W. Chen, and Y. Fang, “Joint channel and queue aware
1947, Sep. 2017. scheduling for latency sensitive mobile edge computing with power
[19] R. Liu, Q. Chen, G. Yu, and G. Y. Li, “Joint user association and resource constraints,” IEEE Trans. Wireless Commun., vol. 19, no. 6, pp. 3938-
allocation for multi-band millimeter-wave heterogeneous networks,” IEEE 3951, Jun. 2020.
Trans. Commun., vol. 67, no. 12, pp. 8502-8516, Dec. 2019. [36] M. Simsek, A. Aijaz, M. Dohler, J. Sachs, and G. Fettweis, “5G-enabled
[20] B. Soleimani and M. Sabbaghian, “Cluster-based resource allocation and tactile internet,” IEEE J. Select. Areas Commun., vol. 34, no. 3, pp. 46-
user association in mmWave femtocell networks,” IEEE Trans. Commun., 473, Mar. 2016.
vol. 68, no. 3, pp. 1746 - 1759, Mar. 2020. [37] M. F. Nuts, “Matrix geometric solutions in stochastic models: An
[21] F. Boccardi et. al., “Spectrum pooling in mmWave networks: Opportu- algorithmic approach,” Courier Corporation, 1981.
nities, challenges, and enablers,” IEEE Commun. Mag., vol. 54, no. 11, [38] M. Serror, C. Dombrowski, K. Wehrle, and J. Gross, “Channel coding
pp. 33-39, Nov. 2016. versus cooperative ARQ: Reducing outage probability in ultra-low latency
[22] H. Shokri-Ghadikolaei, F. Boccardi, C. Fischione, G. Fodor, and M. wireless communications,” in Proc. IEEE Globecom Workshop 2015, pp.
Zorzi, “Spectrum sharing in mmWave cellular networks via cell associa- 1-6, Dec. 2015.
tion, coordination and beamforming,” IEEE J. Sel. Areas Commun., vol. [39] Y. Polyanskiy, H. V. Poor, and S. Verdu, “Channel coding rate in the
34, no. 11, pp. 2902-2917, Nov. 2016. finite blocklength regime,” IEEE Trans. Inf. Theory, vol. 56, no. 5, pp.
[23] M. Rebato, F. Boccardi, M. Mezzavilla, S. Rangan, and M. Zorzi, 2307-2359, May 2010.
“Hybrid spectrum sharing in mmWave cellular networks,” IEEE Trans. [40] H. W. Kuhn, “The Hungarian method for the assignment problem,”
Cogn. Commun. Netw., vol. 3, no. 2, pp. 155-168, Jun. 2017. Naval Res. Logist., vol. 2, nos. 1-2, pp. 83-97, Mar. 1955.
[24] Y. Liu, X. Fang, and M. Xiao, “Discrete power control and trans- [41] R. T. Rockafellar and R. Tyrrel, Network Flows and Monotropic Opti-
mission duration allocation for self-backhauling dense mmWave cellular mization, Wiley, 1984.
networks,” IEEE Trans. Commun., vol. 66, no. 1, pp. 432-447, Jan. 2018. [42] L. R. Ford and D. R. Fulkerson, “Maximal flow through a network,”
[25] R. G. Stephen and R. Zhang, “Joint millimeter-wave fronthaul and Canadian Journal of Mathematics, vol. 8, pp. 399-404, 1956.
OFDMA resource allocation in ultra-dense CRAN,” IEEE Trans. Com- [43] J. Edmonds and R. M. Karp, “Theoretical improvements in algorithmic
mun., vol. 65, no. 3, pp. 1411-1423, Mar. 2017. efficiency for network flow problems,” Journal of the ACM (Association
[26] L. Zhu et. al., “Millimeter-wave NOMA with user grouping, power for Computing Machinery), vol. 19, no. 2, pp. 248-264, 1956.
allocation and hybrid beamforming,” IEEE Trans. Wireless Commun., vol. [44] U. Bauer, “Assignment Problem with Constraints,” Masters thesis, ETH
18, no. 11, pp. 5065-5079, Nov. 2019. Zrich Department of Computer Science, 2004.

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 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/TCOMM.2020.3036046, IEEE
Transactions on Communications

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. XX, NO. X, XX 15

[45] A. H. A. El-Atta and M. I. Moussa, “Student project allocation with Jiantao Yuan received the B.E. degree in electronic
preference lists over (student, project) pairs,” in Proc. 2nd Int. Conf. and information engineering from Dalian University,
Comput. Electric. Eng., Dubai, Dec. 2009, pp. 375-379. Dalian, China, in 2009, the M.S. degree in signal
[46] D. Gale and L. S. Shapley, “College admissions and the stability of and information processing from The First Research
marriage, American Mathematical Monthly, vol. 69, no. 1, pp. 9-15, Jan. Institute of Telecommunications Technology, Shang-
1962. hai, China, in 2012, and the Ph.D. degree from
[47] Y. Liu, X. Fang, M. Xiao, and S. Mumtaz, “Decentralized beam the College of Information Science and Electrical
pair selection in multi-beam millimeter-wave networks, IEEE Trans. Engineer, Zhejiang University, Hangzhou, China. He
Commun., vol. 66, no. 6, pp. 2722-2737, Jun. 2018. was with Datang mobile communication equipment
co. LTD, Shanghai, China, from 2012 to 2013,
where he was involved in LTE network planning and
optimization.
He currently holds a post-doctoral position at the Institute of Ocean Sensing
and Networking of the Ocean College, Zhejiang University, Zhoushan, China.
His research interests include cross-layer protocol design, 5G new-radio
based access to unlicensed spectrum (NR-U), and ultra-reliable low latency
communications (uRLLC).

Rui Liu received the B.S.E. degree from Zhejiang


University, Hangzhou, China, in 2016, where he
is currently working towards the Ph.D degree with
the College of Information Science and Electronic
Engineering. From 2018 to 2020, he was also a
Visiting Student at the School of Electrical and
Computer Engineering, Georgia Institute of Technol-
ogy, Atlanta, GA, USA. His general research inter-
ests include LTE in unlicensed spectrum (LTE-U),
millimeter-wave (mmWave) communications, and
machine learning for wireless networks.
Geoffrey Ye Li (S’93-M’15-SM’97-F’06) Geoffrey
Ye Li is a Chair Professor at Imperial College
London, UK. Before moving to Imperial in 2020, he
was with Georgia Institute of Technology in Georgia,
USA, for twenty years and with AT&T Labs -
Research in New Jersey, USA for five years. His
general research interests include statistical signal
processing and machine learning for wireless com-
munications. In the related areas, he has published
over 500 journal and conference papers in addition
to over 40 granted patents. His publications have
been cited over 40,000 times and he has been recognized as a Highly Cited
Researcher, by Thomson Reuters, almost every year.
Dr. Geoffrey Ye Li was awarded IEEE Fellow for his contributions to signal
Guanding Yu (S’05-M’07-SM’13) received the B.E. processing for wireless communications in 2005. He won several prestigious
and Ph.D. degrees in communication engineering awards from IEEE Signal Processing Society (Donald G. Fink Overview Paper
from Zhejiang University, Hangzhou, China, in 2001 Award in 2017), IEEE Vehicular Technology Society (James Evans Avant
and 2006, respectively. He joined Zhejiang Univer- Garde Award in 2013 and Jack Neubauer Memorial Award in 2014), and IEEE
sity in 2006, and is now a Professor with the College Communications Society (Stephen O. Rice Prize Paper Award in 2013, Award
of Information Science and Electronic Engineering. for Advances in Communication in 2017, and Edwin Howard Armstrong
From 2013 to 2015, he was also a Visiting Professor Achievement Award in 2019). He also received the 2015 Distinguished ECE
at the School of Electrical and Computer Engineer- Faculty Achievement Award from Georgia Tech.
ing, Georgia Institute of Technology, Atlanta, GA, He has been involved in editorial activities for over 20 technical journals,
USA. His research interests include 5G communi- including the founding Editor-in-Chief of IEEE JSAC Special Series on ML
cations and networks, mobile edge computing, and in Communications and Networking. He has organized and chaired many
machine learning for wireless networks. international conferences, including technical program vice-chair of the IEEE
ICC’03, general co-chair of the IEEE GlobalSIP’14, the IEEE VTC’19 (Fall),
Dr. Yu has served as a guest editor of IEEE Communications Magazine and the IEEE SPAWC’20.
special issue on Full-Duplex Communications, an editor of IEEE Journal
on Selected Areas in Communications Series on Green Communications
and Networking, and a lead guest editor of IEEE Wireless Communications
Magazine special issue on LTE in Unlicensed Spectrum, and an Editor of IEEE
Access. He is now serving as an editor of IEEE Transactions on Green Com-
munications and Networking, an associate editor of IEEE Journal on Selected
Areas in Communications Series on Machine Learning in Communications
and Networks, and an editor of IEEE Wireless Communications Letters. He
received the 2016 IEEE ComSoc Asia-Pacific Outstanding Young Researcher
Award. He regularly sits on the technical program committee (TPC) boards
of prominent IEEE conferences such as ICC, GLOBECOM, and VTC. He
has served as a Symposium Co-Chair for IEEE Globecom 2019 and a Track
Chair for IEEE VTC 2019’Fall.

0090-6778 (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: Istanbul Medipol Universitesi. Downloaded on November 12,2020 at 07:32:41 UTC from IEEE Xplore. Restrictions apply.

You might also like