IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO.

1, MARCH 2023 471

Resource Allocation in an Open RAN

System Using Network Slicing
Mojdeh Karbalaee Motalleb , Vahid Shah-Mansouri , Member, IEEE,
Saeedeh Parsaeefard , Senior Member, IEEE, and Onel Luis Alcaraz López , Member, IEEE

Abstract—The next radio access network (RAN) generation, demands high capacity and throughput, e.g., 8K video stream-
open RAN (O-RAN), aims to enable more flexibility and openness, ing and immersive gaming. Meanwhile, the URLLC service
including efficient service slicing, and to lower the opera- provides ultra-reliable and low-latency connectivity, e.g., for
tional costs in 5G and beyond wireless networks. Nevertheless,
strictly satisfying quality-of-service requirements while establish- autonomous vehicles, Tactile Internet, and remote surgeries.
ing priorities and promoting balance between the significantly Finally, mMTC services require connectivity for a large num-
heterogeneous services remains a key research problem. In this ber of Internet of Things (IoT) devices that transmit small
paper, we use network slicing to study the service-aware baseband payloads [4], [5], [6].
resource allocation and virtual network function (VNF) activation
in O-RAN systems. The limited fronthaul capacity and end-to-
end delay constraints are simultaneously considered. Optimizing
A. Motivation
baseband resources includes O-RAN radio unit (O-RU), phys-
ical resource block (PRB) assignment, and power allocation. The optimal resource allocation in 5G systems is crucial
The main problem is a mixed-integer non-linear programming for reducing costs and improving the performance experienced
problem that is non-trivial to solve. Consequently, we break it by the user equipments (UEs). These systems face significant
down into two different steps and propose an iterative algorithm
that finds a near-optimal solution. In the first step, we refor- challenges, including interference alignment, limited capacity
mulate and simplify the problem to find the power allocation, of the fronthaul links, energy restrictions on virtual machines
PRB assignment, and the number of VNFs. In the second step, (VMs), etc. [2], [7], [8].
the O-RU association is resolved. The proposed method is vali- Many studies have investigated resource allocation in cloud
dated via simulations, which achieve a higher data rate and lower RAN (C-RAN) by considering a single service’s power, data
end-to-end delay than existing methods.
rate, and delay limitations. Unfortunately, the existing radio
Index Terms—Open radio access network (O-RAN), virtual access networks (RANs) currently lack adequate flexibility
network function (VNF), network slicing, knapsack problem, and openness to handle these simultaneous service demands.
greedy algorithm, Karush-Kuhn-Tucker (KKT) conditions.
Hence, a new RAN paradigm, called open RAN (O-RAN)
architecture, has emerged. Therefore, O-RAN can simulta-
I. I NTRODUCTION neously support multiple services at a lower cost by being
flexible, layered, and modular. One of the fundamental prob-
ETWORK slicing is a key technology in 5G wireless
N systems. Specifically, it isolates network resources into
slices, e.g., via core slicing and/or radio access network (RAN)
lems lies in balancing services with different QoS, resource
requirements, and priorities in O-RAN architecture [1], [9],
[10], [11].
slicing, for serving various services [1], [2], [3]. The purpose of this paper is to design a system in the
There are three main service classes in 5G, namely O-RAN architecture to support the three types of 5G services,
enhanced mobile broadband (eMBB), ultra-reliable low namely, eMBB, URLLC, and mMTC via network slicing and
latency communications (URLLC), and massive machine-to- resource allocation.
machine communications (mMTC). Each service is assigned
to a network slice depending on its corresponding quality of
service (QoS) requirements. For instance, the eMBB service B. Main Contributions
Manuscript received 24 March 2022; revised 13 July 2022 and 5 September This paper studies the resource utilization of a downlink
2022; accepted 5 September 2022. Date of publication 13 September 2022; O-RAN system to develop an isolated network slicing outline
date of current version 7 March 2023. The associate editor coordinating for the three 5G services. We use mathematical methods to
the review of this article and approving it for publication was J.-F. Botero.
(Corresponding author: Mojdeh Karbalaee Motalleb.) decompose and convexify the problem and solve it using hier-
Mojdeh Karbalaee Motalleb and Vahid Shah-Mansouri are with the archical algorithms. The main contributions of this paper are
School of ECE, University of Tehran, Tehran 1439957131, Iran (e-mail: summarized as follows:
[email protected]; [email protected]@gmail.com).
Saeedeh Parsaeefard is with the Department of Electrical Engineering, • We examine the problem of baseband resource alloca-
University of Toronto, Toronto, ON M5S 3G4, Canada (e-mail: tion, such as power, physical resource blocks (PRBs),
[email protected]). O-RUs, and activating VNFs, to maximize the weighted
Onel Luis Alcaraz López is with the Centre for Wireless Communications,
University of Oulu, 90570 Oulu, Finland (e-mail: [email protected]). throughput of the O-RAN architecture. The three types
Digital Object Identifier 10.1109/TNSM.2022.3205415 of 5G service classes, i.e., eMBB, URLLC, and mMTC,
1932-4537 
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
472 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO. 1, MARCH 2023

TABLE I
together with their corresponding QoS requirements and L IST OF ACRONYMS
service priorities, are considered.
• We propose a two-step resource management algorithm
for solving the optimization problem. In the first step,
we reformulate and simplify the problem so as to find
an upper and lower bound for the number of acti-
vated VNFs. Moreover, we use the Lagrangian function
and Karush-Kuhn-Tucker (KKT) conditions to obtain the
optimal power and PRB allocation. In the second step, the
problem of O-RU association is converted to a multiple
knapsack problem and solved by a greedy algorithm.
• We analyze the complexity of the proposed algorithms
and demonstrate their convergence. Additionally, we ana-
lyze the feasibility region of the problem and introduce
a fast algorithm to check it numerically.
• We show via numerical results that the proposed algo-
rithm outperforms two baseline schemes in terms of
achievable data rate and mean total delay. Remarkably,
the proposed algorithm performs close to the optimal
solution in low-interference conditions.

C. Organization
This paper is organized as follows. Relevant literature
related to our work is discussed in Section II, while Section III
briefly overviews the O-RAN architecture. The system model
and the problem formulation are described in Section IV
and Section V, respectively. The details of our proposed
resource management algorithm are introduced in Section VI.
In Section VII, numerical results are provided to evaluate the to guarantee the QoS requirements of eMBB and URLLC
performance of the proposed algorithm. Finally, Section VIII services. Moreover, [20] investigates the minimization of the
concludes the paper. For clarity, Table I lists the main system’s power for the RAN slicing of eMBB and URLLC
acronyms used throughout the paper. downlink services using non-orthogonal multiple access tech-
niques. In [5], the problem of resource allocation for the
II. R ELATED L ITERATURE coexistence of eMBB and URLLC services is formulated and
solved by deep reinforcement learning.
The network slicing problem in multi-tenant cellular
In [8], [21], VMs activation and beamforming allocation
networks has received significant attention recently, e.g., [7],
are discussed in C-RAN systems. Paper [8] minimizes energy
[12], [13]. Two levels of dynamic network slicing in het-
cost with system delay, fronthaul capacity, and rate constraints.
erogeneous C-RAN (H-CRAN) are examined in [7]. The
To guarantee UE delays, M/M/1 queueing theory is used for
higher level manages user acceptance control, RRH associ-
transmission and processing delays. In [22], [23], the problem
ation, and the allocation of BBU capacity. Meanwhile, PRB
of joint virtual computing resource allocation with beamform-
and power are allocated at lower levels. In [14], RAN slic-
ing is formulated. Also, the association of RRH to the UE is
ing is considered for the fog RAN (F-RAN) system, and
considered and solved using innovative methods.
executed using deep reinforcement learning. In [15], [16],
In [24], [25], the problem of joint power allocation and RRH
the implementation of RAN level slicing is discussed among
association in a H-CRAN system is considered to maximize
multiple mobile network operators with the specific physical
the energy efficiency. Finally, in [26], the optimum power is
network resources infrastructure. In [17], to provide 5G slicing
obtained in the massive MIMO aided C-RAN system, and
services, the authors present a framework called O-RANFed
the problem of RRH to BBU and RRH to UE association is
that implements and optimizes federated learning tasks in
formulated and solved.
O-RAN devices. Moreover, The authors of [18] propose a
federated deep reinforcement learning algorithm to achieve
network slicing in O-RAN. III. BACKGROUND
Recent research has focused on multiplexing eMBB and O-RAN is an appropriate alternative to the next generation
URLLC services within a RAN. In [2], the sum rate of the of radio access networks due to its flexibility, openness, low
eMBB, and URLLC, is optimized by ensuring that each user’s operational costs, and intelligence.
traffic load demand is met and the slice isolation is guaran- O-RAN was developed to jointly benefit from the advan-
teed, thus promoting an amicable service coexistence. In [19], tages of virtual RAN (vRAN) and cloud RAN (C-RAN). By
a RAN slicing is considered in a coordinated multipoint system virtualizing RANs, operators can improve flexibility, reduce

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
MOTALLEB et al.: RESOURCE ALLOCATION IN AN OPEN RAN SYSTEM USING NETWORK SLICING 473

IV. S YSTEM M ODEL

We consider a downlink (DL) system, and an O-RAN archi-
tecture using RAN slicing as depicted in Fig. 1. In this section,
we present the system and signal model, derive the achievable
data rates, power of O-RU, and the fronthaul capacity of the
O-RAN system. Moreover, we discuss the mean delay and the
power of VNFs.

A. System Model
Assume, there are three service types: eMBB, URLLC,
and mMTC, which support different applications. Accordingly,
there are S1 slices for the first service type (eMBB), S2
Fig. 1. Network sliced O-RAN system.
slices for the second service type (URLLC), and S3 slices
for the third service type (mMTC). Therefore, there are
CAPEX and OPEX, and add new capabilities to their networks S = S1 + S2 + S3 pre-allocated slices serving these services.
more quickly. The C-RAN architecture divides the RAN into Moreover, each service request s ∈ {1, . . . , S } is served by
two major parts: the radio remote head (RRH) and the base- its corresponding slice. So we have the set {1, 2, . . . , S1 } of
band unit (BBU). Several distributed RRHs can be connected eMBB service instances, the set {1, 2, . . . , S2 } of URLLC ser-
to a centralized BBU, called BBU-pool [27]. Unlike C-RAN, vice instances, and the set {1, 2, . . . , S3 } of mMTC service
O-RAN separates RAN into three different units, namely instances. Each service sj ∈ {1, 2, . . . , Sj } consists of Usj
Radio Unit (O-RU), Distributed Unit (O-DU), and Central Unit requests from single-antenna UEs requiring certain level of
(O-CU). Mostly non-real-time baseband processing occurs in QoS. Notice that j ∈ {1, 2, 3} indicates the service type. Based
the O-CU layer, while real-time baseband processing occurs on the application and QoS request, UE may be admitted and
in the O-DU layer. allocated to the resources.
In the O-RAN architecture, the PHY is divided into low Each pre-allocated slice contains reserved VNFs for the
and high PHY, unlike C-RAN. As shown in Fig. 1, O-RU is three logical nodes:
• MAC/RLC functions in the O-DU
a logical node that contains RF and low PHY. The former
• PDCP/SDAP functions in the O-CU-UP
transmits or receives radio signals, while the latter includes
• UPF which is a functional layer
digital beamforming. Typically, the O-DU constitutes a logical
node with high PHY, MAC, and RLC. It contains a subfunction Each slice s ∈ {1, 2, . . . , S }, consists of Msd VNFs for the
of the eNodeB and is deployed near the O-RU. Moreover, O- processing of O-DU, Msc VNFs for the processing of O-CU-
DU is connected to an O-RU with an open fronthaul interface. UP, and Msu VNFs for the processing of UPF. The VNFs
In addition to supporting the lower layers of the protocol stack, of O-DU, O-CU-UP, and UPF are interconnected, which is
O-CU also provides support for the higher layers. defined as the service function chain in the O-RAN system.
The O-CU contains two parts: the O-CU user plane (O- Also, each VNF instance runs on a VM that uses resources
CU-UP) and the O-CU control plane (O-CU-CP). The former from the data centers.
hosts the packet data convergence protocol (PDCP)-UP and Assume there are K PRBs in this system. Suppose each slice
the service data adaption protocol (SDAP), while the latter s consists of K̄s pre-allocated virtual resource
 blocks that are
hosts PDCP-CP and radio resource control (RRC). O-DU and mapped to PRBs. Therefore, we have s K̄s ≤ K . In addi-
O-CU are connected via an open and well-defined interface tion, there are R multi-antenna O-RUs that are shared between
F1 . Moreover, O-CU-UP is connected to user plane function the slices. Specifically, the O-RU r ∈ R = {1, 2, . . . , R} has
(UPF) via O-backhaul link. J antennas for transmitting and receiving data. Moreover, all
The O-RAN architecture contains other principal log- O-RUs have access to all PRBs.
ical nodes called Orchestration and Automation, RAN
Intelligent Controller (RIC)- Near Real-Time, and O-Cloud. B. Signal Model
Orchestration and Automation include functions such as RIC Let yu(s,i) be the received signal of UE i in the sth service
Non-Real-Time. RIC is responsible for machine learning such that
methods and making the system more intelligent.  Ks
R 
key feature of the O- RAN architecture is that the hard- yu(s,i) = hH k r k k
r ,u(s,i) gu(s,i) er ,u(s,i) xQ r ,u(s,i) + zu(s,i) ,
ware is disaggregated from the software, leading to network r =1 k =1
function virtualization (NFV). Additionally, each component is (1)
implemented as a virtual network function (VNF), the system where xQ kr ,u(s,i) = xP kr ,u(s,i)
+ q r , xP kr ,u(s,i)
= w kr ,u(s,i)
function block in NFV, that can be deployed on a virtual 
machine (VM) or container [28]. As a result, as shown in prk,u(s,i) xu(s,i) , xu(s,i) depicts the transmitted symbol vec-
Fig. 1, O-RAN components, such as UPF, O-CU, O-DU, and tor, zu(s,i) ∼ CN(0, BN0 ) is the receive additive Gaussian
RIC-near real-time, are virtualized and implemented as VNFs noise, and BN0 is the noise power in a given bandwidth B.
[9], [10], [11], [29], [30], [31], [32]. Here, xP denotes the precoded message before compression,

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
474 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO. 1, MARCH 2023

while xQ illustrates the precoded message after compression. where Δh r ,u(s,i) ∼ N(0, φ 2r ,u(s,i) ) indicates the estimat-
In addition, q r ∼ CN(0, σq 2 IR ) indicates the quantization ing error vector with a Gaussian distribution and φ r ,u(s,i) =
Gaussian noise which comes from the signal compression in diag(φr ,u(s,i) , . . . , φr ,u(s,i) ).
r
O-DU. Furthermore, gu(s,i) ∈ {0, 1} is a binary variable that
illustrates whether O-RU r serves the ith UE that is allocated C. Achievable Data Rate
to the sth slice or not. Furthermore, prk,u(s,i) represents the The achievable data rate for the ith UE request in the s1th
transmission power of the O-RU r serve the ith UE in slice s application of service type 1 (eMBB) can be written as
and PRB k, while hkr ,u(s,i) ∈ CJ corresponding channel vec-

R
tor. In addition, wrk ,u(s,i) ∈ CJ depicts the associated transmit Ru(s1 ,i) = Rr ,u(s1 ,i) gu(s
r
, (5)
1 ,i)
beamforming vector. Therefore, the SINR of the ith UE served r =1
at slice s on PRB k is given by where
prk,u(s,i) |hkr ,u(s,i)
H wrk ,u(s,i) |2 
K
ρkr ,u(s,i) = , (2) Rr ,u(s1 ,i) = Rkr ,u(s1 ,i) erk,u(s1 ,i) (6)
BN0 + Irk,u(s,i)
k =1
A UE in an O-RU r using PRB k receives interference from is the achievable data rate of RU r to UE i in slice s1 , which
other O-RUs in the set R\r that are using the same PRB k. depends on the achievable data rate per PRB, i.e.,
Two types of interference occur between UEs in each slice:
i) inter-slice interference between signals transmitted over dif- Rkr ,u(s1 ,i) = B log2 (1 + ρkr ,u(s1 ,i) ), (7)
ferent slices, and ii) intra-slice interference between signals
transmitted over the same slice. Since the blocklength in URLLC and mMTC is finite, the
Network slicing techniques significantly reduce inter-service achievable data rate for the ith UE request in the application
(inter-slice) interference. One way to leverage a two-time of service type 2 (URLLC) and 3 (mMTC) is not achieved
scale PRB scheduling is to isolate PRBs in slices (in the from the Shannon capacity formula. Instead, in a short packet
first time scale) and schedule the PRBs to the UEs of the transmission, the achievable data rate is approximated as [2]
slices (in the second time scale). Another method consists of 
R
allocating part of the PRBs of eMBB services to URLLC Ru(sj ,i) = Rru(sj ,i) gu(s
r
j ,i)
, (8)
and mMTC [2], [5], [33]. In this paper, we assume that r =1
the PRB scheduling is performed. Also, in Section V-A, where
we briefly study the PRB scheduling between slices. Since
there are limited resources, inter-service interference cannot be Rr ,u(sj ,i) = Rkr ,u(sj ,i) eu(s
k
j ,i)
, (9)
eliminated entirely. Nevertheless, isolating the slices reduces
is the achievable data rate of RU r to UE i in slice s1 , which
inter-service interference considerably and allows us to ignore
depends on the achievable data rate per PRB, i.e.,
it mathematically.
Back to (2), Irk,u(s,i) is the sum of the power of interfering Rkr ,u(sj ,i) = B log2 (1 + ρkr ,u(sj ,i) − ζu(s
k
)e k , (10)
j ,i) u(sj ,i)
signals and quantization noise, and can be represented as
R
 where
Irk,u(s,i) = σq 2 |h kr ,u(s,i) |2 
k
ζu(s j ,i)
= log2 (e)Q −1 () Cku(s ,i) /Nu(s
k . (11)
j =1 j j ,i)
  
(quantization noise) Here,  is the transmission error probability, Q −1 is the inverse
Us R of the Q function, Cku(s ,i) = 1− 1
 k k k
  k 2 depicts the chan-
|hkr H k r 2 j (1+ρu(s )
+ eu(s,i) eu(s,l) pu(s,l) ,u(s,i) wr  ,u(s,l) gu(s,l) | , j ,i)
nel dispersion of UE i at slice sj and PRB k, while Nu(s k
l=1 r  =1 j ,i)
l=i r  =r
   represents the corresponding transmit blocklength. Rr ,u(sj ,i)
(intra-slice interference) is the achievable data rate that is transmitted by O-RU r to
(3) UE i requesting service sj .
k
If we replace pu(s,l) k
and pu(n,l) in (3) by Psmax , an upper
where k is a binary variable that indicates whether the kth
eu(s,i)
bound Īrk,u(s,i) is obtained for Irk,u(s,i) . Therefore, R̄u(s,i) ∀s, i
PRB is allocated to the UE i in slice s, assigned to rth O-RU,
or not. Furthermore, there is no inter-slice interference, only is derived by using Īrk,u(s,i) instead of Irk,u(s,i) in (8) and (5).
intra-slice interference, since slices are assumed to be isolated.
Herein, we consider a zero forcing beamforming vector, D. Power of the O-RU and the Fronthaul Capacity
which minimizes the experienced intra-slice interference, and Let Pr denote the power of the transmitted signal from the
is given by [34] rth O-RU to all the UEs served by it. From (1), the power of
wrk ,u(s,i) = ĥkr ,u(s,i) (ĥkr ,u(s,i)
H
ĥkr ,u(s,i) )−1 . (4) each O-RU r is obtained as follows,

where hkr ,u(s,i) is the channel estimate, which is assumed 

S 
Ks 
Us
Pr = |wrk ,u(s,i) |2 αrk ,u(s,i) + σq2 , (12)
imperfect. Mathematically, ĥ r ,u(s,i) = h r ,u(s,i) +Δh r ,u(s,i) , s=1 k =1 i=1
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
MOTALLEB et al.: RESOURCE ALLOCATION IN AN OPEN RAN SYSTEM USING NETWORK SLICING 475

where αrk ,u(s,i) = prk,u(s,i) gu(s,i)

r erk,u(s,i) . Since we have 1) Processing Delay: Assume the packet arrival of UEs
a fiber link between O-RU and O-DU, the rate of users follows a Poisson process with arrival rate λu(s,i) for the ith
on the fronthual link between O-DU and the rth O-RU is UE of the sth service (or slice). Therefore, the mean
 arrival data
formulated as rate of the sth slice in the UPF layer is αsU = U s
λ
u=1 u(s,i) .
Assume the mean arrival data rate of the UPF layer for

S 
Ks 
Us
slice s (αsU ) is approximately equal to the mean arrival data
Cr = log2 1+ |wrk ,u(s,i) |2 αrk ,u(s,i) /σq2
rate of the O-CU-UP layer (αsC ) and the O-DU (αsD ), i.e.,
s=1 k =1 i=1
αs = αsU ≈ αsC ≈ αsD . This is because the amount of
= log2 Pr /σq2 . (13) data transferred along the route (regardless of frame changes)
is constant. In fact, according to Burke’s theorem, the mean
arrival data rate of the second and third layers, which are
E. Mean Delay processed in the first layer, is still Poisson with rate αs .
In this part, the end-to-end mean delay for each service is It is assumed that there are load balancers in each layer
obtained. The total delay (T tot ), is the sum of the process- for each service to divide equally the incoming traffic to
ing delay (T proc ), the transmission delay (T tr ), and the total VNFs. Suppose the baseband processing of each VNF is
propagation delay (T pro ). modeled by an M/M/1 processing queue. Each packet is
processed by one of the VNFs of the corresponding slice.
T tot = T proc + T tr + T pro , (14a) Therefor, the mean delay for the sth slice in the O-DU, the
T proc = T RU + T DU + T CU + T UPF , (14b) O-CU, and the UPF is modeled as M/M/1 queue, and can be
respectively [8], [22], [23],
T tr = T fr,t + T mid,t + T b,t , (14c)
T pro = T fr,p + T mid,p + T b,p . (14d) TsDU = 1/(μds − αs /Msd ), (15)
TsCU = 1/(μcs − αs /Msc ), (16)
Mathematically, the total propagation delay (T pro ) is the sum of TsUPF = 1/(μus − αs /Msu ), (17)
the propagation delay in the fronthaul link T fr ,p , the midhaul
link T mid,p , and the backhaul link T b,p . In each link, the where Msd , Msc and Msu represent the number of VNFs in O-
propagation delay is the time a signal to reach its destination. DU, O-CU-UP and UPF, respectively. Moreover, 1/μds , 1/μcs ,
It is obtained based on the length of the fiber link and the capacity and 1/μus are the mean service times of the O-DU, O-CU, and
of the link (as T = L / c, where L is the length of the link and c the UPF layers, respectively. The arrival rate of each VNF for
is the propagation speed of the medium). Meanwhile, the total each slice s is αs /Msi i ∈ {d , c, u}.
transmission delay (T tr ) is the sum of the transmission delay On the other hand, arrival data rate  of wireless link for
in the fronthaul T fr ,t , the midhaul T mid,t , and the backhaul each UE i of service s is λu(s,i) , thus U i=1 λu(s,i) = αs .
s

T b,t . In each link, the transmission delay is the amount of time Moreover, the service time of transmission queue for UE i
required to push all the packets into the transmission medium, requesting service s has an exponential distribution with mean
and can formulated as T = R α , where R is the data-rate of 1/Ru(s,i) and can be modeled as a M/M/1 queue [8], [22],
the packet and α is the mean packet size. Notice that taking [23]. Therefore, the mean delay of the transmission layer for
the propagation and transmission delays into account in the UE i in slice s is
formulation is straightforward, but we have avoided it for the RU
Tu(s,i) = 1/(Ru(s,i) − λu(s,i) ). (18)
sake of succinctness and simplicity. Therefore, the propagation
delay is fixed and does not affect the optimization problem. proc
tot
we assume Tu(s,i) ≈ Tu(s,i) .
Next, we present a brief calculation of propagation delay.
Assume a distance between the O-RU and O-DU around 10
F. VNF Power
km, the distance between O-DU and O-CU around 80 km,
not greater than the distance from O-CU to the network Assume the power consumption of each VNF in each logial
around 200 km [35]. Then, assuming the fronthaul, mid- node (O-DU, O-CU, and UPF) in the slice s, is represented by
haul and backhaul are connected with fiber optics and c is φds , φcs , and φus , respectively. Then, the system’s total
 cost of
the speed of like, the propagation delay is about T pro = energy of all the slices can be represented as φtot = Ss=1 φs .
(10 + 80 + 200) × 103 /(3 × 108 ) < 1 ms. A significant issue facing the industry is reducing energy con-
The following is a brief calculation of the transmission sumption. Data centers are one of the most energy-consuming.
delay to show that its contribution to the total delay is As a result, restrictions are placed on data centers’ energy,
negligible and does not affect the optimization. In URLLC including VMs. So, one of our goals is to limit the energy con-
and mMTC, the mean packet size may be between 20 to sumption of total VNFs that can be run as VM on data centers.
32 bytes, while the minimum data rate is assumed to be So, by applying a custom policy on total power consumption, we
1 bps/Hz × BW (180 KHz ). Thus, the transmission delay can control data centers’ power consumption (φtot ≤ φmax ).
from O-RU to O-DU is about T fr ,t = 1×180×10 20×8
3 < 0.1 ms,

and T fr ,t ≈T mid,t ≈ T . For eMBB, the packet size may

b,t V. P ROBLEM S TATEMENT
be 100 times larger and the delay does not exceed 0.6 ms. Suppose the slice s (which is assigned to service s) has a
Therefore, in the following, we assume that the total delay is factor δs (based on the priority of its hosting service)
priority
S
approximate to the processing delay (T tot ≈ T proc ). where s=1 δs = 1. The priority factor of each slice is
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
476 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO. 1, MARCH 2023

obtained according to the service level agreement to promote to the slices. Secondly, we assign PRBs of slices to UEs,
a fairness in the system. This paper aims to maximize the find the optimal number of VNFs for each slice, allocate
sum-rate of all UEs subject to QoS constraints as follows power of UEs, and assign O-RU to UEs, which uses the
proposed Algorithm VI. Suppose, Rmin s , and Rs
max are the

S 
Us
max δs R̄u(s,i) (19a) minimum data rate and maximum data rate of each UE in
P ,E ,M ,G slice s, respectively. Firstly, we need to find the average
s=1 i=1
subject to Pr ≤ Prmax ∀r , (19b) PRB number used by the UEs in each service. Since mMTC
and URLLC require usually short packet transmissions, each
prk,u(s,i) ≥ 0 ∀i , r , s, k , (19c)
UE in mMTC and URLLC requires 1 PRB. So if slice s
prk,u(s,i) ≤ Psmax ∀i , r , s, k , (19d) serves mMTC or URLLC services, with Us UEs, it requires
Ks = Us × 1 PRBs. For eMBB, assume the average rate of
R̄u(s,i) ≥ Rmin
s ∀s, (19e) each UE in slice s serving eMBB UEs is R̄s = B log2 (1+ ρ¯s ),
max
Cr ≤ Cr ∀r , (19f) where, ρ¯s is the average SINR of UEs in slice s. Therefore,
tot max
Tu(s,i) ≤ Ts ∀i , s, (19g) the minimum number of PRBs that slice s with Us UEs
requires is Ksmin = Us × RR̄ s
max . Moreover, the maximum
μs ≥ αs /Ms ∀s, (19h) s
number of PRB that the slice s with Us UEs requires is
R̄u(s,i) ≥ λu(s,i) ∀i , s, (19i)
Ksmax = Us × RR̄min s
. Also, Ks = (Ksmin + Ksmax )/2 is
0 ≤ Ms ≤ Msmax ∀s, (19j) s
the average number of required PRBs in slice s. We aim to
φtot ≤ φmax , (19k) obtain the number of PRBs assigned to each slice s (K¯s ) by

r
gu(s,i) = 1 ∀s, i , (19l) solving
∀r

Ks 
S
r
gu(s,i) erk,u(s,i) ≥ 1 ∀s, i , r , (19m) max δs Ks ln(K¯s ) (20a)
K¯s
k =1 s=1

 S Us subject to K¯s ≤ K , (20b)
r
gu(s,i) erk,u(s,i) ≤ 1 ∀s, i , r , (19n) s
s=1 i=1
r Ksmin ≤ K¯s ≤ Ksmax ∀s ∈ S1 , (20c)
gu(s,i) ∈ {0, 1} ∀s, i , (19o)
K¯s ≤ Ks ∀s ∈ S2 , S3 . (20d)
k
er ,u(s,i) ∈ {0, 1} ∀s, i . (19p)
We use logarithms to assign PRBs to all slices to make them
Here, R̄u(s,i) is derived by using Īrk,u(s,i) instead of Irk,u(s,i) equally fair, since proportional fairness is achieved by maxi-
in (8) and (5). In addition, P = [prk,u(s,i) ], ∀s, i , r , k , is mizing the log utility function [36]. Equation (20b) illustrates
the four-dimensional (4D) matrix of power for UEs, E = that the sum of PRBs of slices can not exceed the maximum
[erk,u(s,i) ], ∀s, i , r , k indicates the binary 4D matrix for the number of PRBs (K). Equation (20c) restricts the number of
PRB association. Moreover, G = [gu(s,i) r ], ∀s, i , r is a PRBs of eMBB slices and (20d) limits the number of the PRBs
binary three dimensional (3D) matrix for the O-RU associ- of URLLC and mMTC slices. By relaxing K¯s , the objective
ation. Furthermore, M = [Msd , Msc , Msu ], ∀s is a matrix function and constraints become convex and can be solved
containing the number of VNFs in each layer of slice. Notice using the Lagrangian function.
that (19b), (19c) and (19d) limit the power of each O-RU
and UE. Also, (19e) constrains the rate of each UE requesting B. Slice Management
each type of service, i.e., eMBB, mMTC, and URLLC, to be In this section, we will look at the life cycle of network
greater than a threshold. Meanwhile, (19f) and (19g) represent slicing on a practical level. The goal is to examine slice
the limited fronthaul capacity and the limited end-to-end delay management, which includes creating, managing, and delet-
of the received signal, respectively. (19h) and (19i) are related ing slices. Network slices generally have four life cycle
to the stability of the M/M/1 queue, (19j) restrictes the number stages [37]:
of VNFs in each slice due to the limited resources, while (19l) • Preparation phase: the operator plans to create a network
and (19m) guarantee that the O-RU and PRB are associated slice instance (NSI) by designing the its template,
with the UE, respectively. Also, (19n) ensures that each PRB onboarding users, and preparing the environment. Also,
can not be assigned to more than one UE associated with the the evaluation of requirements is performed in this step.
same O-RU, (19k) indicates that the fixed cost of energy of • Commissioning phase: the NSI is created, and the
VNFs in each slice does not exceed the threshold, while (19o) requirements are considered and allocated to the slice.
and (19p) constrain E and G to be binary matrices. • Operation phase: the NSIs are activated, managed, moni-
tored (e.g., KPIs), modified, and deactivated. As the slice
A. PRB Scheduling enters the activated phase, it is ready to support services,
In this section, we provide a brief study on the problem and as the slice exits the de-activated phase, the slice is
of PRB scheduling which can be completed in two steps inactive, and communication services are stopped.
to eliminate the inter-slice interference and guarantee the • Decommissioning phase: an NSI that is decommissioned
isolation of slices [36]. Firstly, we should assign the PRBs no longer exists after this phase.
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
MOTALLEB et al.: RESOURCE ALLOCATION IN AN OPEN RAN SYSTEM USING NETWORK SLICING 477

Since the requirements evaluation is considered in the prepa- is obtained by replacing Irk,u(s,i) with Īrk,u(s,i) in (8) and (5)
ration phase, we need an algorithm to estimate the UE traffic and make these equations become concave functions.
in the system at different times. Moreover, based on this esti- Suppose ρ̂kr ,u(s,i) = Psmax |hH
r ,u(s,i) wr ,u(s,i) gu(s,i) | /(BN0 ).
k k r 2
mation, we need to evaluate resources, including the optimal We replace ρkr ,u(s,i) with ρ̂kr ,u(s,i) in (11), to convexify the (8)
number of VNFs, PRB assignment of UEs for each slice, for the URLLC and mMTC services that have the short packet
and the total power requirements. In this phase, we use our transmission. So, a lower bound for (8) is given that is a
algorithm to calculate resources after estimating the system’s concave function.
traffic. As shown in Fig. 1, we have three different slices for
eMBB, URLLC, and mMTC. The system must prepare VNFs Ks
r
j
for MAC/RLC functions in O-DU, PDCP/SDAP functions in R̄u(sj ,i) =B k
eu(s j ,i)
(log2 (1 + ρ̄ku(s2 ,i) ) − ζ̂u(s
k
j ,i)
), (21a)
O-CU, UPF, SMF, and AMF functionality layers for each k =1
slice. Moreover, O-RU, high PHY in O-DU, and O-CU-CP  r
R
are shared between slices. Thus, we do not require evaluat- R̄u(sj ,i) = R̄u(sj ,i) , (21b)
ing and preparing for the share environments and platforms in r =1

the network slicing cycles. Moreover, the estimation of PRB k
ζ̂u(s j ,i)
= log2 (e)Q −1 () Ĉku(s ,i) /Nu(s
k , (21c)
and power is needed based on the proposed algorithm. After j j ,i)

evaluating, assessing, and preparing the resources and environ- Ĉku(sj ,i) = 1 − 1/(1 + ρ̂ku(sj ,i) )2 . (21d)
ments for each slice, the commissioning phase is started. In
this phase, the slices are created based on the previous phase Without loss of generality, assume that UPF, O-CU and O-
estimation. These created slices are activated in the opera- DU use the processors with the same processing capability. We
tion phase, and the actual resources are assigned based on notice that it makes the formulation simpler. However, loosing
the proposed algorithm. It is possible to modify the slice’s this assumption does not change the formulation significantly
resources even when the evaluation changes during the oper- and the problem can be solved in the same manner. Therefore,
ation phase. If we need to remove a slice or any service not we have μs = μus ≈ μcs ≈ μds . Moreover, as mentioned before,
used in a zone, the unshared resources are released in the the mean arrival data rate of the UPF layer for a service s (αsU )
decommissioning phase. is equal to the mean arrival data rate of the O-CU-UP layer
(αsC ) and O-DU (αsD ). So αs = αsU ≈ αsC ≈ αsD . Again,
VI. P ROPOSED A LGORITHM this assumption only simplifies the notations and loosing it
does not make the solution inefficient. These assumptions lead
In this section, we first apply some simplifications to the to having the same processing power for each layer φus =
system; Solving the problem (19) is complicated since this is φcs = φds . As a result, we have Ms = Msu = Msc = Msd .
non-convex mixed-integer non-linear problem (MINLP) with Using the above assumption, we have TsDU = TsCU = TsUPF
a binary variable and an integer variable. We applied some and we have Tsproc = TsRU + TsDU + TsCU + TsUPF . So,
simplifications and use an iterative heuristic algorithm to solve Ts
proc
= TsRU + 3 × TsDU .
the problem. We solve this problem in two levels, iteratively, The problem (19) is mixed-integer nonlinear programming
until it converges [25]. with two integer variables, the PRB assignment, e, and the
At the first level, the main purpose is to assign appro- number of VNFs in slice s, Ms , and by relaxing the variables,
priate PRBs and power to the UEs. Furthermore, sufficient the problem is also non-convex; therefore, this problem is NP-
activated VNFs are assigned to each slice. Hence, at this level, hard. Solving the problem is not trivial. To solve the problem
we would like to obtain the variables P, E, and M. Despite by inspiring Stackelberg, we reformulate the equation in (19g)
the simplification of the problem (19), it is still NP-hard and to reduce one of the variables (i.e., Ms ) that can be solved
challenging to solve. Therefore, we relax the variable E [7], after obtaining the rate of UEs. We notice that Ms is similar to
[25] and reformulating the constraint (19g), to turn them into the followers in Stackelberg Competition, and power and PRB
a jointly-convex problem; Afterward, we solve this problem assignment are identical to the leader. So, the new problem has
using a conventional dual Lagrangian method. In the second two variables: power and PRB assignment. This new problem
level, finding the optimal O-RU association, G, is concerned is convex by relaxing the binary variable, the PRB assignment,
with the fixed parameter of power, PRB allocation, and the and estimating the lower bounds (21). The objective function
number of activated VNFs. We repeat this procedure until the and constraints of the problem are convex and can be solved
algorithm converges. by the Lagrangian function. After obtaining the power of UEs
and PRB assignment, we can obtain the achievable rate of
A. Sub-Problem 1 each UE so we can find the optimal number of VNFs in each
Suppose that G is fixed, we want to obtain P, E and M. slice (Ms ).
Here, we first simplify and relax the parameters to convexify In the following, we define a lemma to find the upper and
k
the problem. As we mentioned before, by replacing pu(s,l) and lower bounds for the optimal number of VNFs based on the
k
pu(n,l) in (3) with Psmax , an upper bound Īrk,u(s,i) is obtained achievable rates. Afterward, we obtain the formula to attain
the optimal number of VNFs.
for Irk,u(s,i) , and also the lower bound ρ̄ku(s,i) is achieved for Lemma 1: The optimal number of VNFs in each slice s can
ρku(s,i) . Moreover, the lower bound R̄u(s,i) , ∀s, ∀i for Ru(s,i) be achieved by the Ms = max{Mu(s,i) |i ∈ 1, 2, . . . , Us } ∀s.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
478 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO. 1, MARCH 2023

αs (T max R
u(s,i) −T max λ
u(s,i) −1)
where, Mu(s,i) = (T maxs μ −3)(R s
for each UE Assume v, m, h, ξ , χ, q and κ are the matrix of
s s u(s,i) −λu(s,i) )−μs
i in slice s. Lagrangian multipliers that have non-zero positive elements.
Proof: In problem (19), the constraint (19g) can be refor- The Lagrangian function is written as
mulated as Us
S 

1 3 L(P, E ; v, χ , h, ξ , κ , m) = δs R̄u(s,i)
Tsmax ≥ + , (22a) s=1 i=1
Ru(s,i) − λu(s,i) μs − αs /Ms
 Us
S  R

αs (Tsmax Ru(s,i) − Tsmax λu(s,i) − 1) + hu(s,i) (R̄u(s,i) − ηu(s,i) ) − mr (Pr − ζr )
Ms ≥ . (22b)
(Tsmax μs − 3)(Ru(s,i) − λu(s,i) ) − μs s=1 i=1 r =1
S  Us  R 
K 
Also from equations in (19k), (19h) and (19j), we have + κkr ,u(s,i) prk,u(s,i) − vkr ,u(s,i) (erk,u(s,i) − 1)
s=1 i=1 k =1 r =1
αs /μs ≤ Ms ≤ min{M max , φmax /3φs }. (23) 
+ ξrk,u(s,i) erk,u(s,i)
+ qkr ,u(s,i) (Psmax − prk,u(s,i) )
We denote Ms = min{M max , φmax /3φs }. Thus, if we R  Us
S 
K 
restrict constraint (19g) to equality, constraint (19g) is still   s

+ χr ,u(s,i) erk,u(s,i) − 1 . (28)

valid. Also, we have the following inequality r =1 s=1 i=1 k =1

αs /μs ≤ Mu(s,i) ≤ Ms , (24) Lemma 2: The derivatives of the Lagrangian function (28)
with respect to the P and E give the KKT conditions to obtain
αs (T max R
u(s,i) −T max λ
u(s,i) −1)
where Mu(s,i) = (T maxs μ −3)(R s
≥ 0 since the the optimal value of these two variables [7], [25].
s s u(s,i) −λu(s,i) )−μs
numerator and the denominator both have the same sign. In Proof: Assume UE i in slice s, associated with O-RU r, is
the numerator, according to (19i), Ru(s,i) − λu(s,i) ≥ 0, and allocated to PRB k (i.e., erk,u(s,i) = 1). Therefore, we have
as we know that αs ≥ 0, we have αs (Ru(s,i) − λu(s,i) ) ≥ the following KKT condition
0. If we assume that the (Ru(s,i) − λu(s,i) )Tsmax ≥ 1, the ∂L
numerator will be positive. (Ru(s,i) −λu(s,i) )Tsmax ≥ 1 since = (δs + hu(s,i) )Bkr ,u(s,i)
∂prk,u(s,i)
the order of Tsmax is about milli second and the difference
between achievable rate and packet rate can be more than + (skr ,u(s,i) − Dkr ,u(s,i) ) = 0, (29)
1/Tsmax . Therefore, to ensure that this constraint will be valid,
we restrict constraint (19i) to Ru(s,i) ≥ λu(s,i) + 1/Tsmax . So where skr ,u(s,i) = κkr ,u(s,i) − qkr ,u(s,i) and other parameters
the numerator will be positive. In the denominator, we can say are as follows
that (Tsmax μs )(Ru(s,i) − λu(s,i) ) − μs ≥ 0, since, μs ≥ 0 Dkr ,u(s,i) = mr |wrk ,u(s,i) |2 gu(s,i)
r
erk,u(s,i) , (30a)
and (Ru(s,i) − λu(s,i) ) ≥ 1/Tsmax as mentioned above. The
Bkr ,u(s,i) = B |hH k k 2 r k k
r ,u(s,i) wr ,u(s,i) | gu(s,i) er ,u(s,i) Sr ,u(s,i) / ln(2),
left side of the equation (24), leads to Ru(s,i) ≥ λu(s,i) that
is the constraint (19i). For the right side, by reformulating the (30b)
equation (24), we have a new constraint ∀i , ∀s given by Skr ,u(s,i) = (|hH k k 2 k
r ,u(s,i) wr ,u(s,i) | kr ,u(s,i) + BN0 + Irk,u(s,i) )−1 .

Ru(s,i) ≥ λu(s,i) + 3/(Tsmax μs − αs Tsmax /Ms − 3) (30c)

+ 1/Tsmax = u(s,i) . (25) Also, kkr ,u(s,i) = gu(s,i)
r erk,u(s,i) prk,u(s,i) . Thus, from equa-
tion (29), optimal power is obtained and power is allocated.
Thus, to obtain the optimal number of activated VNF in each
We denote jkr ,u(s,i) = gu(s,i)
r erk,u(s,i) . The optimal power is
slice, we need to find the maximum of the Mu(s,i) in each slice
as Ms = max{Mu(s,i) |i ∈ 1, 2, . . . , Us } ∀s. as follow.
Despite simplifying the problem in (19), it is still non- prk,u(s,i)
convex and hard to solve. Therefore, the conventional approach  +
(δs + hu(s,i) )B jkr ,u(s,i) BN0 + Irk,u(s,i)
to solve the problem of the PRB and the power allocation = − .
ln(2)(Dkr ,u(s,i) − skr ,u(s,i) ) |hH k wk |2 jkr ,u(s,i)
is to relax the variable E into continuous value erk,u(s,i) ∈ r ,u(s,i) r ,u(s,i)

[0, 1] ∀s, ∀i , ∀r , ∀k [7], [25]. Furthermore, the problem can be (31)

solved using the Lagrangian function and iterative algorithm. Also [a]+ = max(0, a). In addition, PRB assignment can be
In order to transform (19) into a convex optimization achieved from the derivatives of the Lagrangian function (28)
problem in standard form, it is required to change the vari- with respect to the E as follow.
r
able of equations (13) to Pr = σq2r × 2C so the constraint
r ∂L
(19f) is changed to Pr ≤ σq2r × 2Cmax . The combination of = R̄rk,u(s,i) (δs + hu(s,i) ) − mr |wrk ,u(s,i) |2 prk,u(s,i) gu(s,i)
r
∂erk,u(s,i)
equations (19b) and (19f) leads to the following equation
r + (ξrk,u(s,i) − vkr ,u(s,i) + χr ,u(s,i) ) = 0. (32)
P r ≤ ζr = min{Pmax , σq2r ×2 Cmax
}. (26)
So, the optimal E is obtained using the KKT conditions, which
Moreover, the combination of equations in (19e), (19i) require solving
and (25) leads to the following equation
erk,u(s,i) (Fkr ,u(s,i) − vkr ,u(s,i) − mr |wrk ,u(s,i) |2 prk,u(s,i) gu(s,i)
r
) = 0,
1
R̄u(s,i) ≥ ηu(s,i) = max{Rmin
u(s,i) , λu(s,i) + s , u(s,i) }. (27) (33)
Tmax

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
MOTALLEB et al.: RESOURCE ALLOCATION IN AN OPEN RAN SYSTEM USING NETWORK SLICING 479

Algorithm 1 Greedy Algorithm for Assignment of O-RU to Algorithm 2 Iterative Algorithm for the Baseband Resource
UEs (GAA) Allocation and VNF Activation (IABV)
u(s,i) = R ∀s, ∀i , ∀r . 1: Set the maximum num. of iter. Imax , convergence condition >
r
1: Set gu(s,i) = 0, Cr = tr , and Brem
2: Sort slices according to their δs in descending order 0
3: for s ← 1 to S do 2: Assign Users to O-RU randomly (Initialize G)
4: for i ← 1 to Us do 3: for i ← 1 to Imax do
5: RU = 0 4: Acquire P (i) , E (i) and M (i) using Lagrangian function
6: for r ← 1 to R do and sub-gradient method based on (VI-A)
r
7: Acquire Gru(s,i) = R̄u(s,i) 5: Update G (i) based on algorithm GAAOU (1) in (VI-B)
6: if the algorithm converged with the tolerence of then
8: end for
Break
Obtain r ∗ = argmaxr ∈Brem
7:
9: Gru(s,i)
u (s,i) 8: else: Continue the algorithm
10: while RU == 0 do 9: end if
11: if Cr ∗ ≥ ψr ∗ ,u(s,i) then 10: end for
r ∗
12: Set gu(s,i) =1
13: Set Cr ∗ = Cr ∗ − ψr ∗ ,u(s,i)
14: Set RU = 1
∗ obtained (27) in Section VI-A, we can ignore this constraint
15: else: Brem rem
u(s,i) = Bu(s,i) \ {r }
in (35). The problem (35) is an NP-complete 0-1 multiple
16: end if knapsack problem. We solve this problem using the heuris-
17: end while
18: end for tic GAA Algorithm 1, which is a greedy algorithm [7], [38].
19: end for
r
Firstly, we set all the variables to zero (gu(s,i) = 0, ∀s, i , r ).
Then we define the parameter Bu(s,i) . This parameter is used
rem

as a set of O-RUs that can be assigned to the UE i in slice

where Fkr ,u(s,i) = R̄rk,u(s,i) (δs + hu(s,i) ) + (ξrk,u(s,i) + s, which initially includes all the O-RUs (Brem u(s,i) = R, ∀s, i ).
χr ,u(s,i) ). Hence, from equation (32) and (33), PRB assign- Also we introduce another parameter Cr = tr , ∀r which is the
ment is performed as follows knapsack capacity of each O-RU. Next, we sort all the slices
 based on their priority. Afterward, based on the sorting of the
1 u(s, i ) = argmax Zkr ,u(s,i) ∀r , k ∈ K , s ∈ S ,
erk,u(s,i) = (34) UEs, we assign the O-RU that provides the highest achievable
0 otherwise,
data rate for each UE on the condition that the value of the
where Zkr ,u(s,i) = (Fkr ,u(s,i) − vkr ,u(s,i) − mr |wrk ,u(s,i) |2 desired UE (ψr ,u(s,i) ) does not exceed the knapsack capacity
of each O-RU (Cr ). If it exceeds the capacity of the desired
prk,u(s,i) gu(s,i)
r ).
O-RU, we remove the specific O-RU from the set of O-RUs
Thus, the user in slice s that has the most considerable value ∗
that can be assigned to that UE (Brem u(s,i) = Bu(s,i) \ {r }).
rem
of Fkr ,u(s,i) , should be allocated to PRB k. Since just one
Then, the O-RU with the highest achievable data rate from
PRB can be allocated to a UE between those UEs (regardless
the new set of O-RUs Brem u(s,i) is selected. The complexity
of the services), that is associated
  toS the same O-RU. The of sorting S  slices based on their priority is O(S log(S )).
number of UEs are N = Ss=1 U i=1 1. Also, assume that
the algorithm converges after Tconv times. The complexity Depict N = Ss=1 U s
i=1 1 as the whole number of UEs in
order of this problem is about O(Tconv × N × K ). the system. The complexity order of this algorithm is about
O(S log(S )) + O(R × N).
B. Sub-Problem 2
After power allocation and PRB assignment, the remaining C. Iterative Proposed Algorithm
problem is to assign O-RU to each UE in each service. For In Sections VI-A and VI-B, the details of solving each sub-
this, assume P and E are fixed, so we want to find G. Next, problem are depicted. Here, the iterative algorithm for the
we introduce a greedy algorithm that assigns an O-RU to each whole problem is demonstrated. Firstly, we fixed G to achieve
UE. P and E, using the Lagrangian method and the KKT condi-
Greedy Algorithm Assignment (GAA): The following is a tions. Afterward, G is updated using the GAA algorithm. This
reformulated version of the problem. process is repeated until it converges. The whole algorithm

S 
Us 
R (IABV method) is depicted as follows (Algorithm 2).
r
max r
δs gu(s,i) R̄u(s,k ) (35a) 1) Complexity Order: The number of UEs are N =
G S
s=1 i=1 r =1 s=1 US . Also, assume that the algorithm converges after
S  Us Tconv times. As we mentioned before, the complexity order
subjectto r
gu(s,i) ψr ,u(s,i) ≤ tr ∀r , (35b) of the first sub-problem is about O(Tconv × N × K ) and
s=1 i=1 the complexity order of the second sub-problem is about

r
gu(s,i) = 1 ∀s, i , (35c) O(S log(S )) + O(R × N). So the complexity of the main
r problem (19) is O(Tconv × N × K × (S log(S ) + RN)).
r
gu(s,i) ∈ {0, 1} ∀s, i , (35d) 2) Convergence Analysis: Due to limited system resources,
K s we have limits on VNFs’ power, UE or O-RU power, fronthaul
where ψr ,u(s,i) = k =1 |wr ,u(s,i) | pr ,u(s,i) er ,u(s,i) and
k 2 k k
capacity, etc. As a result, the objective function, which is
tr = ζr − σr because of the equations (26) and (12). Since we the aggregate throughput, cannot exceed its optimal value
Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
480 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO. 1, MARCH 2023

and become infinite. Therefore, if the aggregate through-

put is infinite and increases without limit, the resources
must also be unlimited. Hence, the system has an optimal
solution: its maximum aggregate throughput in the feasible
region.
Consequently, we can guarantee the convergence of the
iterative algorithm if the objective function is the ascending
function concerning the number of iterations and has an upper
bound. Thus, it will converge to its optimum value if it is a
strictly ascending function and to its local optimum if it is a
non-monotonically ascending function.
Us the aggregate throughput as T(P, E, G) =
Fig. 2. O-RU placement in a cell.
Consider
S
s=1 i=1 δs R̄u(s,i) . In the first step of the iteration i of
the Algorithm 2 (IABV), we have T(Pi , Ei , Gi−1 ). In this VII. N UMERICAL R ESULTS
step, optimal power and PRB allocation are obtained for
the fixed O-RU association, so we have T(Pi , Ei , Gi−1 ) ≥ In this section, firstly, we describe the initial points and the
T(Pi−1 , Ei−1 , Gi−1 ). In the second step of the iteration i, the comparison algorithms. Then, we talk about the feasible region
optimal O-RU association is achieved to maximize the aggre- of our system model. Afterward, we illustrate the numerical
gate throughput. So we have this inequality T(Pi , Ei , Gi ) ≥ results.
T(Pi , Ei , Gi−1 ). As a result, we have T(Pi , Ei , Gi ) ≥
T(Pi−1 , Ei−1 , Gi−1 ). Hence, in each step of the iteration, the A. The Initial Points and the Comparison Algorithms
aggregate throughput increased. Note that T∗ (P∗ , E∗ , G∗ ) is In this part, numerical results for the main problem are
the achieved aggregate throughput for all the feasible resource depicted to evaluate the performance of the algorithms using
allocation solutions of {P, E, G}. So, T∗ (P∗ , E∗ , G∗ ) ≥ the Monte-Carlo method. We consider three network slices for
T(Pi , Ei , Gi ) and thus in each iteration, the aggregate eMBB, URLLC, and mMTC services. Assume we have six 4-
throughput can not be larger than the optimal solution. antenna O-RU (MISO) located in a place with a diameter of
In addition, if we assume that the interference is set to be 500 meters as shown in Fig. 2. In addition, we consider the
zero Irk,u(s,i) = 0, and we suppose that each UE has the maxi- users placed randomly in this area.
mum power prk,u(s,i) = Psmax , and we consider that all PRB is Here, the channel vector from the O-RU r to the UE i in ser-
L L
vice s is set as hkr ,u(s,i) = dr−,u(s,i) Ωkr ,u(s,i) , where dr−,u(s,i)
assigned to all UE erk,u(s,i) = 1∀s, ∀i and each UE is assigned
is the distance between the O-RU r and UE i in service s
to the nearest O-RU with the best channel quality. So, the solu-
and L = 3.8 is the path-loss exponent. Also, Ωkr ,u(s,i) is the
tion of this allocation, is the upper bound for the aggregate
throughput. Thus, we can guarantee the local convergence of random variable that is generated by the Rayleigh distribution
our iterative algorithm since the objective function T is the and it is the Rayleigh fading channel between the UE and O-
ascending function concerning the number of iterations and it RU. We consider 25 PRBs in the network. The packet size
has the upper bound which is not infinite. for mMTC is equal to 20 bytes, and for URLLC is equal to
In addition, to extend our solution to the global optimum, we 32 bytes [39]. The maximum number of VNF for each slice
must prove that the algorithm monotonically increases in the is 25 and the mean arrival data rate for the eMBB service
non-optimal set of solutions and is Lipschitz monotone con- is λ = 3Mbps and for the mMTC service and the URLLC
traction mapping. Here, we briefly discussed our algorithm’s service is λ = 0.2Mbps. Also, the quantization noise is
global convergence for a low interference system. assumed to be 10−13 . Moreover, we set hu(s,i) = ηu(s,i) /200,
In the low interference system, the PRB and VNF assign- mr = ζr /10 and qkr ,u(s,i) = Psmax /100. The other parame-
ment is obtained straightforwardly, and the first step for solving ters of these simulations are depicted in Table II [39], [40],
the problem is power allocation. Given fixed O-RU asso- [41], [42].
ciation, we have T(Pi , Ei , Gi−1 ) > T(Pi−1 , Ei−1 , Gi−1 ) Finding a feasible initial value is almost tricky. We use a fast
(strictly increase). Notice that for the power allocation problem, method discussed in Section VII-B to overcome this challenge.
the objective function is convex, and the convex functions Two different methods are used to compare the performance
are Lipschitz monotone contractions. However, in the second of the proposed method (IABV) and show the optimality of
step of the algorithm, we can show as before that the algo- our approach. The first one is a baseline scheme, which uses
rithm is increased, but we can not talk about monotonically random PRB allocation. Therefore, the allocation of PRB to
increases. Hence we have T(Pi , Ei , Gi ) ≥ T(Pi , Ei , Gi−1 ). each UE is random when we have low interference, but in
Nevertheless, the objective function is still Lipschitz mono- figures with high interference, we randomly assign just one
tone mapping. Accordingly, we have T(Pi , Ei , Gi ) > RB to each UE. Also, the association of O-RU is carried
T(Pi−1 , Ei−1 , Gi−1 ). Therefore the objective function is the out based on distance. It means that each UE is assigned to
strictly ascending function concerning the number of iterations the nearest O-RU. The optimal power is obtained using the
and has an upper bound. Consequently, the algorithm converges CVX of MATLAB, which uses the successive convex approx-
to the global optimum solution in low interference. imation (SCA) method since the problem is convex. After

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
MOTALLEB et al.: RESOURCE ALLOCATION IN AN OPEN RAN SYSTEM USING NETWORK SLICING 481

TABLE II
S IMULATION PARAMETERS

Fig. 3. Aggr. throughput vs. number of UEs in each service.

B. Feasible Region
Applying the correct initial point is crucial in making the
system feasible and convergent. Therefore, we investigated the
Algorithm 3 Fast Algorithm (FA) to Check Convergence non-converging and converging simulation for models with
1: Set count = 0 fixed initial parameters and UE random channel gains. We
2: Set prk,u(s,i) = 0, erk,u(s,i) = 0 and gu(s,i)
r = 0 ∀r , k , s, i experimentally found that in non-converged simulations, there
3: for s ← 1 to S do are UEs at the edge of the boundaries or far away from the O-
4: for i ← 1 to Us do RU and have a weak channel gain. One solution is to eliminate
5: count = count +1 UEs who undermine system convergence. For a large number
6: r ∗ = arg minr dr ,u(s,i) ∀r
r∗
of UEs with a fixed number of PRBs, the probability of hav-
7: gu(s,i) =1 ing an infeasible solution increases due to a large number of
8: temp =mod(count, K) UE interference. Another solution is to remove the simula-
9: if temp=0 then
tions in the Monte-Carlo that do not converge using the fast
10: erK∗ ,u(s,i) = 1
algorithm (FA) to check the convergence before the proposed
11: Set prK∗ ,u(s,i) = min{Psmax , Prmax /N} algorithm (IABV). Therefore, if more than half of the iter-
temp
12: else: er ∗ ,u(s,i) = 1 ∀r ations have a feasible solution for the initial condition, the
13:
temp
Set pr ∗ ,u(s,i) = min{Psmax , Prmax /N} simulation can be displayed as a feasible model. If the condi-
tions in (27), (26), (19d) and (19c) are met in the fast algorithm
14: end if (FA), the given algorithm
15: end for S will Us
converge. Assume, the number
16: end for
of UEs is N = s=1 i=1 1, the number of PRBs is K,
and the distance between the rth O-RU to the UE i in slice s
is dr ,u(s,i) . The FA algorithm is represented in Algorithm 3.
The complexity order of this algorithm is O(R × N) which
achieving power and other parameters, the achievable rate will is remarkably lower than the complexity order of the IABV
be obtained, and the optimal number of VNF is achieved from method. In the FA algorithm, the O-RU association is based
Lemma (1). on the distance of the UE to the O-RU. Each UE is associated
For the second one, we use the idea of the fixed BBU with the nearest O-RU. Also, the power of each UE is set to
capacity and dynamic resource allocation (FBDR) algorithm be the minimum of the maximum power of each UE and the
proposed in [7] and named it the dynamic resource alloca- maximum power of each O-RU divided by the total number
tion scheme (DR scheme). We have services with different of UEs (min{Psmax , Prmax /N}). Moreover, the allocation of
QoS in this work, similar to tenants with different QoS intro- PRBs to UEs is based on dividing the number of UEs by the
duced in [7]. Therefore, we can use the DR scheme similar total number of PRBs.
to the FBDR method adapted to our conditions for compar-
ison. Instead of BBU in C-RAN, we have O-DU and O-CU
in O-RAN. Since we do not talk about O-DU and O-CU C. Performance Results
capacity, we use the dynamic resource allocation scheme (DR In Fig. 3, the aggregate throughput is plotted versus the
scheme) algorithm and do not consider BBU capacity. In the number of UEs in each service for these three methods.
DR scheme, PRB and power are dynamically allocated. The Suppose we have one service instance for each type of service,
number of VNFs is obtained from the simulation. The UEs so we have three various services in this figure. Also, we have
are associated with the O-RU based on the quality of their between 6 to 48 UEs in the system. Here, we did not consider
channels and the channel distance instead of using the greedy the priority. The figure presented that the proposed method,
Algorithm 1 (GAA algorithm) for O-RU assignment. The fig- IABV, is 18.6% higher throughput than the baseline scheme.
ures in [7] show that dynamic BBU capacity and dynamic As the number of UEs increases in each service, the aggregated
resource allocation (DBDR) perform better than FBDR for throughput initially increases. Still, due to the interference and
the same priority area. Notice that our proposed algorithm the power constraint, it will be saturated from 12 UEs in each
performs better than the DR scheme. service.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
482 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO. 1, MARCH 2023

Fig. 4. Number of VNFs in each service vs. arrival rate. Fig. 7. Total Delay of a URLLC vs. arrival rate of a UE.

Fig. 5. Aggr. throughput for eMBB vs. maximum power.

Fig. 8. Aggr. throughput vs. the number of mMTC UEs.

Fig. 6. Total Delay of a URLLC vs. the arrival rate of a UE. Fig. 9. Throughput of eMBB services vs. priority of the first one.

Fig. 4 depicts the number of activated VNFs for five differ- mean delay of a URLLC service does not reach the maximum
ent mean service times of one URLLC service vs. the mean threshold of the delay.
arrival time for 12 UEs. This figure presents that as the mean Fig. 7 is the same as Fig. 6 that presented the mean total
arrival rate increases, the number of activated VNF increases. delay of a UE in a URLLC service regarding the mean arrival
Moreover, the number of activated VNFs decreases when the rate of the UE for 20 UEs using three different methods. As
mean service rate increases. you can see, the proposed method (IABV) outperforms the
In Fig. 5, the aggregate throughput is depicted vs. the max- other scenarios.
imum power of UE for three different instances of eMBB Fig. 8, represents the aggregate throughput concerning the
service using proposed method (IABV), DR scheme and the number of UEs in each service and the maximum power for
baseline scheme. Here, we suppose that we have 12 UEs three different mMTC service instances. Assume each UE in
in each service. We assume that these three services require each mMTC service instance requires 0.1 bits/sec/Hz data rate
5bits/sec/Hz, 10bits/sec/Hz, and 15bits/sec/Hz. As you can and is not sensitive to the end-to-end delay. There is no restric-
see in the figure, increasing the maximum power increases tion on fronthaul link capacity and the number of VNFs. The
the aggregate throughput. Moreover, the proposed method figure depicts that by increasing the number of UEs in each
(IABV), gives higher aggregate rates in compared to the DR instance of the service, or by increasing the maximum power
scheme and the baseline scheme. of each UE in each instance of mMTC service, the aggregate
Fig. 6 illustrates the mean total delay of a UE in a URLLC throughput increases.
service regarding the mean arrival rate of the UE and the num- Assume we have two types of eMBB service instances. In
ber of UEs in the service for the proposed method (IABV). It Fig. 9, the aggregate throughput (by considering the prior-
is shown that the delay is an ascending function of the mean ity factor δs ) is depicted for two eMBB service instances.
arrival rate (when the mean service time is fixed) and the Here we consider 4 UEs in each service. The Fig. 9 presented
number of UEs in the service. Moreover, we can see that the that by increasing the priority factor for one service instance,

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
MOTALLEB et al.: RESOURCE ALLOCATION IN AN OPEN RAN SYSTEM USING NETWORK SLICING 483

Fig. 10. Aggregate throughput vs. number of iterations. Fig. 12. Aggregate throughput vs. number of UEs.

Fig. 11. Mean Delay time vs. number of iterations. Fig. 13. Aggregate throughput vs. maximum interference.

more resources are allocated to this service instance, and the

aggregate throughput of this service is increased and vice
versa. Also, we can realize from this figure that the aggregate
throughput has the most significant value at the same priority.
In Fig. 10, the aggregate throughput is shown according to
the number of iteration (outer loop) of the proposed algorithm
(IABV) for different numbers of UEs for one service. In this
figure, the convergence of the IABV method is illustrated. The
minimum data rate for each UE is assumed to be 2 Mbps. After Fig. 14. Aggregate throughput vs. number of UEs.
four iterations, IABV converges to the fixed value.
In Fig. 11, the mean total delay of URLLC service is indi-
cated according to the number of iterations of the proposed been added to the system. In -105 dBm, we have 5 UEs, and
algorithm (IABV) for different numbers of UEs for one at the end, we have 55 UEs in the system. Since the amount of
URLLC service. This figure shows that the algorithm con- interference in the system is entered as a fixed value, the allo-
verged to the fixed value after four iterations. cation of PRBs is not considered. The higher maximum power
In Fig. 12, the aggregate throughput is shown according threshold leads to a greater aggregate throughput. The aggre-
to the number of UEs for two different methods, namely gate throughput first increases with the number of UEs and at
the proposed algorithm (IABV) and the optimal method for the same time the amount of the maximum interference, then
URLLC service for the low interference. The minimum data it becomes almost fixed and finally decreases so much. When
rate is 5bits/sec/Hz for each UE and the maximum delay is the aggregate throughput decreases, the maximum interference
0.1ms. Also the mean arrival rate is set to be 0.2Mbps and is so high that it takes the system out of feasibility.
the mean service rate is 0.5Mbps. The optimal approach is In Fig. 14, the aggregate throughput is shown versus the
obtained from the two-step joint exhaustive search and using number of UEs for an eMBB service with low interference
CVX. In each iteration in the first step, the PRB allocation for the IABV and FA methods in the feasible region. The
and O-RU association are obtained from brute force, and in minimum data rate for each UE is 1Mb/s/Hz. The maximum
the second step, we use CVX to get optimal power. Our power for each O-RU is 34dBm, and the maximum power
solution is close to the optimal value in a small number of for each UE is 30dBm. We assume that the system is not
UEs. Therefore, this figure numerically demonstrates that the sensitive to fronhaul capacity and end-to-end delay and has
proposed algorithm converges to the global optimal in low enough VNF resources. By increasing the number of UEs, the
interference. aggregate throughput raises. And we can see that the IABV
In Fig. 13, the aggregate throughput is depicted vs. the maxi- method is better than the FA method.
mum interference for two different maximum power thresholds Table III shows the execution time given a number of UEs
of O-RU. Here we assume that with the increase of every ten for one service for the three methods. We run our simulation on
dBm of interference power, it is assumed that ten users have the system with configuration (RAM = 8 GB, CPU = Core

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
484 IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, VOL. 20, NO. 1, MARCH 2023

TABLE III
E XECUTION T IME VS . N UMBER OF UE S [5] M. Alsenwi, N. H. Tran, M. Bennis, S. R. Pandey, A. K. Bairagi, and
C. S. Hong, “Intelligent resource slicing for eMBB and URLLC coexis-
tence in 5G and beyond: A deep reinforcement learning based approach,”
IEEE Trans. Wireless Commun., vol. 20, no. 7, pp. 4585–4600, Jul.
2021.
[6] S. Riolo, D. Panno, and L. Miuccio, “Modeling and analysis of tagged
preamble transmissions in random access procedure for mMTC scenar-
ios,” IEEE Trans. Wireless Commun., vol. 20, no. 7, pp. 4296–4312,
Jul. 2021.
[7] Y. L. Lee, J. Loo, T. C. Chuah, and L.-C. Wang, “Dynamic network slic-
ing for multitenant heterogeneous cloud radio access networks,” IEEE
Trans. Wireless Commun., vol. 17, no. 4, pp. 2146–2161, Apr. 2018.
[8] J. Tang, W. P. Tay, T. Q. S. Quek, and B. Liang, “System cost
i5, SSD Hard Disk). As the number of UEs in the system minimization in cloud RAN with limited fronthaul capacity,” IEEE
increases, the execution time increases polynomially for all Trans. Wireless Commun., vol. 16, no. 5, pp. 3371–3384, May 2017.
three algorithms. Since the baseline scheme is a simpler algo- [9] O-RAN Working Group 1, O-RAN Architecture Description V06.00,
O-RAN Alliance, Alfter, Germany, Mar. 2022.
rithm, with random PRB allocation and O-RU association [10] L. Gavrilovska, V. Rakovic, and D. Denkovski, “From cloud RAN to
based on distance, the execution time is less than the two other open RAN,” Wireless Pers. Commun., vol. 113, no. 3, pp. 1523–1539,
algorithms. Power and PRB are allocated in the DR scheme, Mar. 2020.
[11] N. Kazemifard and V.Shah-Mansouri, “Minimum delay function place-
but O-RUs are associated based on distance. Therefore the ment and resource allocation for Open RAN (O-RAN) 5G networks,”
execution time is less than the proposed algorithm. Comput. Netw., vol. 188, Apr. 2021, Art. no. 107809.
[12] L. Feng, Y. Zi, W. Li, F. Zhou, P. Yu, and M. Kadoch, “Dynamic
resource allocation with RAN slicing and scheduling for uRLLC and
eMBB hybrid services,” IEEE Access, vol. 8, pp. 34538–34551, 2020.
VIII. C ONCLUSION [13] Y. Loong Lee, J. Loo, and T. C. Chuah, “A new network slicing frame-
In this paper, we modeled the downlink of the O-RAN work for multi-tenant heterogeneous cloud radio access networks,” in
Proc. Int. Conf. Adv. Electr. Electron. Syst. Eng. (ICAEES), Nov. 2016,
system using isolated network slicing for different 5G services, pp. 414–420.
i.e., eMBB, mMTC, and URLLC. Our goal is to maximize [14] H. Xiang, S. Yan, and M. Peng, “A realization of fog-RAN slicing via
aggregate throughput by activating VNFs in each service, deep reinforcement learning,” IEEE Trans. Wireless Commun., vol. 19,
no. 4, pp. 2515–2527, Apr. 2020.
associating RUs, allocating power, and PRBs. The limited [15] S. E. Elayoubi, S. B. Jemaa, Z. Altman, and A. Galindo-Serrano, “5G
fronthaul capacity and the mean delay for each service are RAN slicing for verticals: Enablers and challenges,” IEEE Commun.
considered. The problem is mixed-integer non-linear program- Mag., vol. 57, no. 1, pp. 28–34, Jan. 2019.
[16] S. D’Oro, F. Restuccia, and T. Melodia, “Toward operator-to-waveform
ming that is solved by a two-step iterative algorithm. In the 5G radio access network slicing,” IEEE Commun. Mag., vol. 58, no. 4,
first step, we reformulated the problem to achieve the num- pp. 18–23, Apr. 2020.
ber of activated VNFs as a function of data rate. Then, we [17] A. K. Singh and K. K. Nguyen, “Joint selection of local trainers and
resource allocation for federated learning in open RAN intelligent con-
obtained PRB association and power allocation using the trollers,” in Proc. IEEE Wireless Commun. Netw. Conf. (WCNC), 2022,
Lagrangian method. In the second step, the O-RU associa- pp. 1874–1879.
tion is carried out. The performance of our proposed method [18] H. Zhang, H. Zhou, and M. Erol-Kantarci, “Federated deep rein-
(i.e., IABV) is compared with the baseline scheme and DR forcement learning for resource allocation in O-RAN slicing,” 2022,
arXiv:2208.01736.
scheme in [7]. In addition, the feasible region is discussed, [19] P. Yang, X. Xi, T. Q. S. Quek, J. Chen, X. Cao, and D. Wu, “How
and the FA algorithm is introduced to check the feasibility should I orchestrate resources of my slices for bursty URLLC ser-
of the initial values. Also, we assume distinct scenarios for vice provision?” IEEE Trans. Commun., vol. 69, no. 2, pp. 1134–1146,
Feb. 2021.
each service, i.e., eMBB, mMTC, and URLLC, based on their [20] F. Saggese, M. Moretti, P. Popovski, “Power minimization of down-
requirement QoS. Simulation results show that the proposed link spectrum slicing for eMBB and URLLC users,” Jun. 2021,
method (i.e., IABV) achieves 18.6% higher data rate than arXiv:2106.08847.
[21] K. Guo, M. Sheng, J. Tang, T. Q. S. Quek, and Z. Qiu, “Exploiting
the baseline scheme. Moreover, simulation results illustrate hybrid clustering and computation provisioning for green C-RAN,” IEEE
more minor delays for the proposed method (IABV) than DR J. Sel. Areas Commun., vol. 34, no. 12, pp. 4063–4076, Dec. 2016.
scheme and the baseline scheme. [22] P. Luong, F. Gagnon, C. Despins, and L.-N. Tran, “Joint virtual
computing and radio resource allocation in limited fronthaul green C-
RANs,” IEEE Trans. Wireless Commun., vol. 17, no. 4, pp. 2602–2617,
Apr. 2018.
R EFERENCES [23] P. Luong, F. Gagnon, C. Despins, and L.-N. Tran, “A novel energy-
efficient resource allocation approach in limited fronthaul virtualized
[1] X. Shen et al., “AI-assisted network-slicing based next-generation C-RANs,” in Proc. IEEE 87th Veh. Technol. Conf. (VTC Spring), Jun.
wireless networks,” IEEE Open J. Veh. Technol., vol. 1, pp. 45–66, 2018, pp. 1–6.
Jan. 2020. [24] S. Ali, A. Khan, and A. Ahmad, “Energy-efficient resource alloca-
[2] M. Setayesh, S. Bahrami, and V. W. S. Wong, “Joint PRB and power tion and RRH association in multitier 5G H-CRANs,” Trans. Emerg.
allocation for slicing eMBB and URLLC services in 5G C-RAN,” in Telecommun. Technol., vol. 30, no. 1, Jan. 2019, Art. no. e3521.
Proc. IEEE Global Commun. Conf. (GLOBECOM), Taipei, Taiwan, [25] S. Ali, A. Khan, and A. Ahmad, “Joint RRH-association, sub-channel
Dec. 2020, pp. 1–6. assignment and power allocation in multi-tier 5G C-RANs,” IEEE
[3] P. Popovski, K. F. Trillingsgaard, O. Simeone, and G. Durisi, Access, vol. 6, pp. 34393–34402, 2018.
“5G wireless network slicing for eMBB, URLLC, and mMTC: A [26] N. Amani, S. Parsaeefard, H. Taheri, and H. Pedram, “Power-efficient
communication-theoretic view,” IEEE Access, vol. 6, pp. 55765–55779, resource allocation in massive MIMO aided cloud RANs,” Aug. 2019,
2018. arXiv:1908.07568.
[4] A. Dogra, R. K. Jha, and S. Jain, “A survey on beyond 5G network [27] B. Han et al., “Research on resource migration based on novel RRH-
with the advent of 6G: Architecture and emerging technologies,” IEEE BBU mapping in cloud radio access network for HSR scenarios,” IEEE
Access, vol. 9, pp. 67512–67547, 2021. Access, vol. 7, pp. 108542–108550, 2019.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.
MOTALLEB et al.: RESOURCE ALLOCATION IN AN OPEN RAN SYSTEM USING NETWORK SLICING 485

[28] R. Mijumbi, J. Serrat, J.-L. Gorricho, N. Bouten, F. De Turck, Vahid Shah-Mansouri (Member, IEEE) received
and R. Boutaba, “Network function virtualization: State-of-the-art and the B.Sc. degree in electrical engineering from the
research challenges,” IEEE Commun. Surveys Tuts., vol. 18, no. 1, University of Tehran, Tehran, Iran, in 2003, the
pp. 236–262, 1st Quart., 2016. M.Sc. degree in electrical engineering from the
[29] S. Niknam et al., “Intelligent O-RAN for beyond 5G and 6G wireless Sharif University of Technology, Tehran, in 2005,
networks,” May 2020, arXiv:2005.08374. and the Ph.D. degree from The University of British
[30] C. B. Both et al., “System intelligence for UAV-based mission critical Columbia, Vancouver, BC, Canada, in 2011. Since
with challenging 5G/b5G connectivity,” Feb. 2021, arXiv:2102.02318. 2013, he has been with the School of Electrical and
[31] “O-RAN Working Group 2: AI/ML workflow description and require- Computer Engineering, University of Tehran. His
ments,” O-RAN Alliance, Alfter, Germany, Rep. O-RAN.WG2.AIML- research interests include analysis and mathematical
v01.02.02, Mar. 2020. modeling of communication and computer networks.
[32] B. S. Lin, “Toward an AI-Enabled O-RAN-based and SDN/NFV-driven
5G and IoT network era,” Netw. Commun. Technol., vol. 6, no. 1,
pp. 6–15, Jun. 2021.
[33] J. Mei, X. Wang, K. Zheng, G. Boudreau, A. B. Sediq, and H. Abou-
Zeid, “Intelligent radio access network slicing for service provisioning in Saeedeh Parsaeefard (Senior Member, IEEE)
6G: A hierarchical deep reinforcement learning approach,” IEEE Trans. received the B.Sc. and M.Sc. degrees from
Commun., vol. 69, no. 9, pp. 6063–6078, Sep. 2021. the Amirkabir University of Technology (Tehran
[34] S. Huang, H. Yin, J. Wu, and V. C. M. Leung, “User selection Polytechnic), Tehran, Iran, in 2003 and 2006,
for multiuser MIMO downlink with zero-forcing beamforming,” IEEE respectively, and the Ph.D. degree in electrical
Trans. Veh. Technol., vol. 62, no. 7, pp. 3084–3097, Sep. 2013. and computer engineering from Tarbiat Modares
[35] J. Cavazos. “5G testing: What is O-RAN? Part 2.” 2020. [Online]. University, Tehran, in 2012. She is currently a
Available: https://blogs.keysight.com/blogs/inds.entry.html part of Apple Cellular Platform Architecture Team.
[36] D. Marabissi and R.Fantacci, “Highly flexible RAN slicing approach She was a Research Scientist and a Lecturer with
to manage isolation, priority, efficiency,” IEEE Access, vol. 7, the University of Toronto, Canada, from 2019 to
pp. 97130–97142, 2019. 2021. She was a Postdoctoral Research Fellow with
[37] 5G Management and Orchestration; Concepts, Use Cases and the Telecommunication and Signal Processing Laboratory, Department of
Requirements (3GPP TS 28.530 Version 15.0.0, Release 15), ETSI Electrical and Computer Engineering, McGill University, Canada, in 2013.
Standard TS-128-530-V15.0.0, Oct. 2018. From 2010 to 2011, she was a visiting Ph.D. student with the Department of
[38] Y. Akçay, H. Li, and S. H. Xu, “Greedy algorithm for the general Electrical Engineering, University of California at Los Angeles, Los Angeles,
multidimensional knapsack problem,” Ann. Oper. Res., vol. 150, no. 1, CA, USA. Her works are published in three books and 50 journals and con-
pp. 17–29, Dec. 2007. ference papers, which are cited more than 1300 times with an H-index of
[39] “5G; study on scenarios and requirements for next generation access 19 and i10-index of 31. Her background includes optimization theory, game
technologies (3GPP TR 38.913 version 14.3.0 release 14),” ETSI, Sophia theory, signal processing, and networking. Her current research focuses on
Antipolis, France, Rep. TR-138-913-V14.3.0, Oct. 2017. distributed/federate learning, machine-learning applications in next-generation
[40] Evolved Universal Terrestrial Radio Access (E-UTRA); Base Station wireless networks, such as network automation and resource management, and
(BS) Radio Transmission and Reception (Release=13), 3GPP-TS- robustness in both learning and communication. She was awarded the IEEE
36.104-V13.3.0, Mar. 2016. Women in Engineering Award in Region 8.
[41] Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Frequency
(RF) Requirements for LTE Pico Node B (Release 13), 3GPP, Sophia
Antipolis, France, Rep. TR-36.931-V13.0.0, Jan. 2016.
[42] E. Mohyeldin, “Minimum technical performance requirements for IMT
radio interface(s),” ITU, Geneva, Switzerland, Rep. ITU-R M.2410-0, Onel Luis Alcaraz López (Member, IEEE) received
2020. the B.Sc. degree (First-Class Hons.) in electrical
engineering from the Central University of Las
Villas, Cuba, in 2013, the M.Sc. degree in elec-
trical engineering from the Federal University of
Paraná, Brazil, in 2017, and the D.Sc. degree
(with Distinction) in electrical engineering from the
University of Oulu, Finland, in 2020. From 2013 to
2015, he served as a Specialist in Telematics with the
Cuban Telecommunications Company (ETECSA).
Mojdeh Karbalaee Motalleb (Karbalaeimotaleb) He currently holds an Assistant Professorship
received the B.S. and M.S. degrees in electri- (tenure track) of Sustainable Wireless Communications Engineering with
cal engineering from the Amirkabir University of the Centre for Wireless Communications, Oulu, Finland. He is the coau-
Technology, Tehran, Iran, in 2015 and 2017, respec- thor of the book titled Wireless RF Energy Transfer in the Massive IoT
tively. She is currently pursuing the Ph.D. degree Era: Towards Sustainable Zero-Energy Networks (Wiley, December 2021).
with the Department of Computer and Electrical His research interests include sustainable IoT, energy harvesting, wireless RF
Engineering, University of Tehran, Tehran. She is energy transfer, wireless connectivity, and machine-type communications. He
also a visiting research student with the Center for is a collaborator to the 2016 Research Award given by the Cuban Academy of
Wireless Communications, Oulu University, Oulu, Sciences, a co-recipient of the 2019 IEEE EuCNC Best Student Paper Award,
Finland. Her research interests include optimization the author of the Best Doctoral Thesis in Engineering in Finland in 2020, and
and machine learning in network and wireless the recipient of the 2022 Young Researcher Award in the field of technology
system. in Finland.

Authorized licensed use limited to: Universidad Federal de Pernambuco. Downloaded on June 07,2023 at 15:03:23 UTC from IEEE Xplore. Restrictions apply.