You are on page 1of 9

Ain Shams Engineering Journal 12 (2021) 487–495

Contents lists available at ScienceDirect

Ain Shams Engineering Journal


journal homepage: www.sciencedirect.com

Electrical Engineering

A novel scheduling technique for improving cell-edge performance in


4G/5G systems
Wael S. Afifi a,b,⇑, Ali A. El-Moursy c, Mohamed Saad c, Salwa M. Nassar a, Hadia M. El-Hennawy d
a
Computer and Systems Department, Electronics Research Institute, Cairo, Egypt
b
Faculty of Engineering, Ain Shams University, Egypt
c
Department of Computer Engineering, University of Sharjah, Sharjah, United Arab Emirates
d
Department of Electronics and Electrical Communication Engineering, Faculty of Engineering, Ain Shams University, Cairo, Egypt

a r t i c l e i n f o a b s t r a c t

Article history: In cellular networks, users near the edge of the cell are usually suffering from low signal-to-noise-plus-
Received 18 September 2019 interference-ratio (SINR) levels as a result from being far away from the base-station (BS). Many factors
Revised 3 March 2020 could lead to huge attenuation of the received signal in the cell-edge area such as path-loss and multipath
Accepted 29 July 2020
fading. Increasing the BS transmit power is not always feasible as this could lead to an increased inter-
Available online 6 October 2020
cell-interference (ICI). Hence, the cell-edge problem arises. In this paper, a new scheduling technique
has been developed to increase the probability of assigning the available resource blocks (RBs) to the
Keywords:
cell-edge users so that their achieved throughput would increase. A performance comparison with
Wireless networks
Cellular system
state-of-the-art schedulers indicates that our proposed scheduling mechanism leads to a significant
5G improvement in the average throughput for cell-edge users, with negligible performance degradation
LTE schedulers for cell-center users.
Cell-edge Ó 2020 The Authors. Published by Elsevier B.V. on behalf of Faculty of Engineering, Ain Shams University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-
nd/4.0/).

1. Introduction the corresponding service data rates needed to assure a fair perfor-
mance among all users’ requirements. Users located at the edge of
Since the development of mobile wireless communication sys- the cell, i.e., far away from the base station (BS), typically experi-
tems in the late 1970s, huge efforts have been devoted to improve ence a low signal-to-interference-plus-noise-ratio (SINR), which
both the capacity of the system, and the quality-of-service (QoS) leads to considerably low achievable data rates. This is known as
experienced by the users. To achieve these goals, many approaches the cell-edge problem. Increasing the BS transmit power in an
have been taken such as well-planned networks, massive antenna attempt to improve the cell-edge users’ experience is typically lim-
configurations and Inter Cell Interference Coordination (ICIC) ited by the resulting amount of inter-cell-interference (ICI)
methods. In conventional cellular networks, the full area of cover- towards adjacent cells.
age is composed of multiple cells, where each cell is assigned a pre- Most specifications of wireless communication systems have
determined amount of radio resources and power allocation levels. been using a frequency reuse of unity due to the shortage of avail-
Accordingly, the overall limited network resources will be dis- able frequency resources (see, e.g., [1]). Frequency reuse may lead
tributed among the cells based on the capacity of each cell, and to ICI, especially at the cell boundaries, further negatively impact-
ing the throughput of cell-edge users. 5G systems are expected to
be based on an ultra-high densification of base stations to improve
⇑ Corresponding author. the overall system capacity [2]. This densification, however,
E-mail addresses: waelsafifi17@eri.sci.eg (W.S. Afifi), aelmoursy@sharjah.ac.ae together with frequency reuse and high traffic loads, will further
(A.A. El-Moursy), msaad@sharjah.ac.ae (M. Saad), salwa@eri.sci.eg (S.M. Nassar), negatively affect the throughput of cell-edge users. This motivates
helhennawy@ieee.org (H.M. El-Hennawy).
us to study methods for improving the cell-edge user throughput
Peer review under responsibility of Ain Shams University.
for 4G and 5G systems.
LTE [3] and LTE-Advanced (LTE-A) [4,5] are the leading systems
for 4G as stated by the 3rd Generation Partnership Project (3GPP)
[6]. Both LTE and LTE-A use orthogonal frequency division multiple
Production and hosting by Elsevier
access (OFDMA) in the downlink. OFDMA divides the available

https://doi.org/10.1016/j.asej.2020.07.022
2090-4479/Ó 2020 The Authors. Published by Elsevier B.V. on behalf of Faculty of Engineering, Ain Shams University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

transmission bandwidth into multiple sub-carriers, which can The authors of [15] consider allocating RBs to the users based
carry independent data flows. A physical resource block (RB) is on weighted signal-to-noise (SNR) ratios for both cell-centric
defined as 12 subcarriers in the frequency domain (180 kHz) and and cell-edge users, then allocating transmit powers, as two sep-
seven OFDM symbols in the time domain, which is equivalent to arate steps. Their approach relies on combinatorial optimization
one time slot (0.5 ms). RBs are periodically allocated by the sched- and graph-theoretic techniques, which add to the complexity of
uler to the active users in a cell every transmission time interval the system. In contrast, this paper introduces a low-complexity,
(TTI) that is equal to 1 ms. probabilistic approach for the dynamic allocation of RBs among
This paper focuses on improving cell-edge user throughput via cell-centric and cell-edge users. A detailed description for the pro-
intelligently scheduling RBs among the users of the cell. The rest posed approach is illustrated in the next section. Throughout our
of the paper is organized as follows: state-of-the-art techniques experiments, our proposal has proven to achieve larger improve-
for RB scheduling are reviewed in Section 2. In Section 3, The pro- ments as compared to [15], as will be illustrated in the results
posed scheduling technique is explained. In Section 4, Simulation section.
parameters and setup are demonstrated. Simulation experiments In [16], a dynamic PF scheduler has been proposed to increase
and performance results are discussed in Section 5. Finally, Sec- the performance of the cell-edge scheduler. The authors of [16]
tion 6 concludes the paper. modified in the PF scheduler itself to adapt to the individual user
conditions (e.g., location in the cell or QoS requirement). The pro-
posed modification aims to give a chance to the cell-edge users to
2. Related work utilize some of the RBs seeking improvement of their ability to get
the service. This is done by varying the exponent of the dominator
The related literature can be classified into three categories: (1) of the scheduling equation used in conventional PF (i.e., bparame-
the main scheduling techniques (in LTE and LTE-A), (2) scheduling ter) to be not fixed to unity as conventional PF use. The dominator
schemes tailored towards improving cell-edge user throughput is representing the past achieved average throughput by the user
and (3) work on cell-edge throughput improvement via other prior to the processed RB. The algorithm is adapting bsuch that
methods (not scheduling). In what follows we review the main the users near cell center will be discredited since their btends to
related literature from each category. be larger than 1, while the users in the cell edge will be compen-
The main scheduling techniques for LTE and LTE-A can be sum- sated since their btends to be smaller than 1. The dynamic PF algo-
marized as follows (more details can be found in [7,8]). rithm uses a time counter to record the duration in which a user
stays in the cell-edge area. Hence, if a user stays in the cell-edge
 Largest Weighted Delay First (LWDF) scheduler ensures that for a relatively long time, the user will be then prohibited from
each packet has to be received within a certain deadline to being favored by additional RBs. We will demonstrate in the results
avoid packet drops. The scheduler collects information about section that our algorithm will outperform this dynamic PF’s
the creation time of a specific packet, as well as its deadline. performance.
The user with the most stringent requirements in terms of Studies on improving cell-edge user throughput using other
acceptable loss rate and deadline expiration will be preferred methods (than scheduling) do exist. For example, Coordinated
for allocation. This scheduler does not take the channel quality Multi-Point transmission and reception (CoMP) was introduced
variation into consideration. in LTE-A [4,5] and refers to a family of functionalities involving
 Blind Equal Throughput (BET) attempts to achieve throughput multiple BSs coordinating transmissions to a cell-edge user,
fairness among all users. In particular, the scheduler allocates including joint processing and coordinated scheduling. The stud-
resources to flows that have been served with lowest average ies in [17–20] focus on improving cell-edge throughput via CoMP.
throughput in the past. Again, this scheduler is unaware of However, CoMP comes at the cost of coordination and communi-
the channel quality variation. cation across multiple cells. Therefore, this paper focuses on cell-
 Maximum Throughput (MT) scheduler assigns RB to the user edge user improvement via low-complexity uncoordinated tech-
that achieves the maximum throughput. This is usually per- niques that can be readily implemented within the cell. Further-
formed by selecting the user with the largest channel quality more, cell-edge user throughput can also be improved via power
indicator (CQI) for each RB assignment. The major disadvantage control. Typically, power control is applied as a separate step
is that the scheduler does not consider fairness among users. after RB scheduling (see, [15]). Therefore, our introduced schedul-
 Proportional Fair (PF) scheduler strikes a balance between fair- ing mechanism can further benefit from intelligent power alloca-
ness and throughput, by taking both CQI and resource allocation tion performed separately. This paper, however, focuses on
history. In particular, the user with largest ratio of CQI and aver- scheduling.
age past throughput is selected for each RB assignment. Improving cell-edge performance could be implicitly achieved
using techniques enhancing the overall cell throughput. In addition
Heuristics optimization algorithms, such as [9–11], are rarely to the COMP and power control techniques, the smart scheduler
used in LTE resource scheduling [12,13] because of the time com- introduced by NOKIA in [17] is capable to improve cell-edge data
plexity needed in such algorithms. RB scheduling and allocation in rate as well as other users using the following techniques:
LTE is a real-time operation which has to be performed each TTI
(i.e., 1 ms), which is not the case in heuristics optimization algo-  Frequency Selective Scheduling (FSS): FSS uses Channel Aware
rithms that assume plenty of computation power and time to find Scheduling (CAS) and Interference Aware Scheduling (IAS) to
the solution as in offline applications. Works done in [12,13] use select non-faded RBs to each user. Information about fading
very low number of RBs (i.e., maximum one quarter the number channels can be obtained from CQI reports.
of the used RBs in our paper) to bypass the delay caused from used  Modifying the parameters of handover between adjacent cells
heuristics. Due to its fast response, and consideration of both fair- based on the information exchange of the X2 interface. This
ness and throughput, PF becomes the most supported scheduling technique is known as ‘‘Intra-frequency load balancing” and
technique for LTE system [14]. None of the above-mentioned used when the load in two adjacent cells is not balanced.
schedulers, however, pays explicit attention to cell-edge users.  Interference shaping to avoid ICI.
Scheduling techniques that are particularly devoted to cell-  QoS differentiation by allocating more re-sources for users in
edge user throughput enhancement do exist. See, e.g., [15–17]. weak channel conditions.
488
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

Our proposed scheduler can further benefit from any of the the proposed scheduling technique, as to strike a balance
above techniques used in NOKIA scheduler to add to the overall between featured users (who gain more privileges in terms
scheduling. This paper, however, focuses on scheduling. In light of throughput and extra RBs) and non-featured users (who
of the above, the contribution of this paper can be summarized sacrifice some resources).
as follows.
The proposed technique uses two-level scheduling. In the first
 A new scheduling algorithm is proposed to enhance the level, PF scheduling is applied to identify a PF user to potentially
throughput achieved by cell-edge users. The proposed algo- utilize the RB. Then, a second scheduling level will probabilisti-
rithm is an extension to the pro-portional fair scheduler. In par- cally decide whether to allocate that RB to the chosen PF user,
ticular, it is based on probabilistically replacing a cell-centric or to one of the FUs. In contrast to conventional PF scheduling
user by a cell-edge user at some RBs, where the allocation prob- [24] that selects the candidate user to utilize RB based on the
ability changes dynamically across RBs according to the concept ratio between the current instantaneous and the accumulated
of sampling without replacement to give a chance to the cell- average throughput by the user, our algorithm involves putting
edge users for scheduling and utilizing some RBs. However, this the selected PF user into a second round of competition for the
probability should decay gracefully to avoid poor system uti- RB with all FUs (cell-edge users who probably experience weak
lization if significant RBs are assigned to poor signal strengths signal strength). The selection between PF and FU is probabilistic,
cell-edge users. where the selection probabilities vary dynamically as to strike a
 Our proposed scheduler has a low complexity, and is suitable to balance between improving cell-edge user throughput and mini-
be invoked every TTI. mizing the reduction in overall system throughput. To describe
the details of our proposed algorithms, we first make the follow-
Our proposed scheduler aims to improve the cell-edge through- ing definitions.
put, while maintaining the required throughput for cell-centric
users to continue achieving an acceptable level of performance.  The set of FUs in the cell is denoted as {FU1, FU2, . . ., FUN},
where N is the number of FUs in the cell.
 The set of available RBs in the cell per TTI is denoted as {RB1,
3. Edge User Friendly Scheduler (EUFS) RB2, . . ., RBM}, where M is the number of RBs in the cell.
 The minimum required probability of selecting the PF candidate
A new packet scheduling technique is developed to enhance the for RB1 is denoted as P(PF). This is a design parameter that
performance of cell-edge users, or more generally, users experienc- should be identified by the network operator. The effect of the
ing bad channel conditions. The proposed algorithm is an exten- choice of P(PF) on the overall system performance will be
sion to the proportional fair scheduler. In particular, it is based explored later.
on probabilistically replacing a cell-centric user by a cell-edge user  The initial number of chances for the PF user at RB1 is denoted
at some RBs, where the allocation probability changes dynamically as PF_C.
across RBs to give a chance to the cell-edge users for scheduling  The initial number of chances for FUi at RB1 is denoted as FU_Ci.
and utilizing some RBs. However, this probability should decay  Hence, the total number of chances at RB1 could be denoted as:
gracefully to avoid poor system utilization if significant RBs are
assigned to cell-edge users experiencing poor signal strengths. X
N
UE C ¼ PF C þ FU C i
Network operators could offer this feature to the all cell-edge n¼1
users, or to a selected subset of cell-edge users according to some
criteria (such as in return for additional fees). Accordingly, at each
TTI, the core network controller classifies a user as featured user We assume that each FU has the same number K of initial
(FU), i.e., a user that can potentially benefit from the proposed chances of being selected for the first RB (i.e., RB1). In other words,
scheduler, by checking two conditions. These conditions can be initially,
summarized as follows. FU C i ¼ K; for i ¼ 1; 2; . . . ; N: ð1Þ

 FU classification criteria: Guidelines for choosing the scaling factor K as to achieve the best
1. The user has been classified as a cell-edge user: (Obligatory) performance is discussed in Section 5.
Classifying users between cell-centric and cell-edge is very Therefore, and since the probability of selecting a user is calcu-
critical. A user is considered as cell-edge if its instantaneous lated by dividing the number of chances for that user by the total
throughput is below some threshold. This threshold value is number of chances, it is straightforward to see that the initial prob-
chosen to be the 5th percentile point of the overall cell ability of selecting the PF candidate for the first RB (i.e., RB1) is
throughput, as stated in [21]. CQI reports are used to mea- given by PF_C/(PF_C + NK).
sure the instantaneous throughput of a user. Those users Since the minimum required initial probability of selecting a PF
only will be eligible for requesting improved service. user (for RB1) is P(PF), the following condition on the number of
2. The user is agreeing on paying additional fees for the improved chances needs to be satisfied:
service: (Optional) ðPF CÞ
Operators could decide to associate the level of system per- >¼ PðPFÞ ð2Þ
ðPF C þ NKÞ
formance provided to users with pricing, as done in [22,23]
based on the recording of RBs utilization. Nowadays, most The latter implies that PF_C, the initial number of chances for
of the operators are using fixed price charging models that the PF user at RB1, will be set according to:
set constant rental fees for users per time or per consumed
ðPðPFÞ  N  KÞ
bit rate. This works well for systems that include scheduling PF C >¼ ð3Þ
ð1  PðPFÞÞ
techniques implementing blind fairness to users/connections
without any prioritization towards users suffering from For the first RB, a user is selected among the PF and FU candi-
degraded channel conditions (e.g., cell-edge users). Hence, dates depending on their respective selection chances initialized
operators could use different pricing schemes to cope with using (1) and (3), respectively. For every other RB, the selection
489
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

chances are updated according to the statistical concept of sam- Table 1


pling without replacement (SWR). In particular, the user chosen Simulation parameters.

in one RB will be either assigned lower chance/ probability, or even Parameter Value
totally removed from competing with others for the next RB, and System Bandwidth 20 MHz
so on. Consequently, users not selected in a RB will have higher Carrier Frequency 2.14 GHz
chances in the next RBs. If all chances ðPF Cand all FU C i Þreach Channel Model Fast Fading
zero, the chance values PF Cand all FU C i will be reset according Total No. of RBs per TTI 100
Antenna per RRH 3
to (1) and (3), respectively. The process continues until all RBs eNodeB Transmit Power 40 W
are utilized by users. No. of eNodeBs 7
The overall scheduling algorithm (for each cell, at the beginning Total No. of active UEs 210 to 1260
of each TTI) can be summarized as follows. User Mobility No Mobility
No. of Cells 21
No. of TTIs 100
Algorithm EUFS

1. Among all UEs in the cell, identify the featured set {FU1,
FU2, . . ., FUN} based on the FU classification criteria. [16] which have been discussed in the ‘‘related work” section.
2. Set the initial chances PF Cand FU C i ðfor1 <¼ i <¼ NÞac- Throughout our experiments, the proposed algorithm (i.e., ‘‘EUFS”)
cording to (1) and (3), respectively. is deeply investigated and compared to the best recently proposed
3. For each RB1, RB2, . . ., RBM, do: algorithms. It is easily seen that algorithm EUFS is based on
a. Invoke the PF scheduler to select a PF user. dynamically changing the chances (probabilities) of selecting PF
b. Create the set U of users containing PF Ccopies of the and FU candidates, respectively, across the different consecutive
PF user, and FU C i copies of FU i ðfor1 <¼ i <¼ NÞ. RBs. The first experiment investigates the optimal value of the scal-
c. Select one user from U according to a uniform distribu- ing factor K that achieves the best performance using EUFS algo-
tion, to utilize the RB. rithm. The purpose of the second experiment is to assess the
d. Update the selection chances as follows: proposed algorithm EUFS against the state-of-the-art techniques
 If the selected user is a PF user, then PF C :¼ PF C  1; from [15,16]. The throughput achieved by cell-edge and cell-
 If the selected user is FU i ðforany1 <¼ i <¼ NÞ, then centric users is used as the comparison metric. The third experi-
FU C i :¼ FU C i  1. ment also assesses the performance the proposed algorithm EUFS
e. If all chances ðforPFandallFUsÞreach zero, reset PF Cand against the state-of-the-art techniques from [15,16]. The perfor-
FU C i ðfor1 <¼ i <¼ NÞaccording to (1) and (3), mance metric used, however, is the signal-to-noise-ratio (SNR).
respectively. In the fourth experiment, we use Jain’s fairness index [27] to assess
the fairness of the system across all scheduling algorithms. The
concept of Jain’s fairness index is close to the variance measure
in statistical analysis, as it measures the deviation of the achieved
This algorithm ensures allowing the users who previously suf- throughput by each user from the fair throughput. Jain’s fairness
fered from lower instantaneous throughputs (namely cell-edge index is maximum (i.e., unity) when all users achieve the same
users) to utilize additional RBs to improve their throughputs and throughput. The closer the users’ throughputs are to each other,
avoid throttling them. The stolen RBs from PF users are substituted the larger Jain’s fairness index could be attained by the system.
gradually from other users in the same cell who already have bet- In the fifth (last) experiment, the execution time is compared
ter throughput in a dynamic manner that would not affect their across all scheduling algorithms to investigate the speed of each,
assigned bandwidth seriously as well. This purpose has been main- and make sure that the proposed algorithm could fit in LTE online
tained and proven as will be shown in the results and analysis RB scheduling and allocation.
section.
5. Results and analysis
4. Experimental setup
5.1. Assessing the initial scaling factor
The proposed algorithm has been deployed into the Vienna LTE
system-level simulator [25,24], which runs on top of MATLAB. The The first study is performed on the choice of the best value for
simulation parameters used in our experiments are shown in the initial number of chances (K) as used the EUFS algorithm. Our
Table 1. The used system bandwidth and carrier frequency is sup- empirical approach is based on experimenting with a wide range of
ported in LTE Release-8 [26]. The remaining parameters have been values for K, then choose the best performing value for the remain-
selected to be compatible with the system bandwidth used and to ing experiments of the paper. The selection of K will depend on the
be applicable in the realistic world as well. Without loss of gener- performance of the cell-edge users; the effect of EUFS algorithm on
ality, we assume that all cell-edge users are considered as FUs. If cell-centric users, however, is discussed in details in the following
some of the cell-edge users refuse the condition of the additional experiments. Fig. 1 shows the average throughput of cell-edge
pricing to be granted the improved performance, this would imply users using EUFS relative to their average throughputs obtained
that the number of featured users will be decreased. Hence, the from a static probability approach, (on the y-axis) using different
average throughput for the other users in both categories (cell- values for K (on the x-axis). According to (1), K represents the ini-
centric and cell-edge) will be even better than the results provided tial chances assigned to each cell-edge user per cell (i.e., FU C i ).
in next section. However, to keep the initial probability for PF user P(PF) fixed
For performance evaluation, three state-of-the-art algorithms while changing K, scaling K directly indicates a corresponding scal-
are used as benchmarks to be compared against our proposal. ing for the PF user (i.e., PF_C) accordingly. At the first RB0, P(PF) is
The first algorithm is the conventional PF algorithm which is the only effective factor in assigning RB0 to either Cell-edge or cell-
adopted by the standard LTE system [14]. The other two algorithms centric user. However, the choice of K would take effect while step-
are the weighted SNR algorithm [15] and the dynamic PF algorithm ping up in the following RBs as the chances adaptation after each
490
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

Fig. 1. Average throughput for cell-edge users using EUFS relative to the average throughput for cell-edge users using fixed probabilities, across different K.

RB would be dissimilar. Small K values (i.e., K  number of RBs) Now, the average throughput of all considered cell-edge users
indicate an abrupt change to the total number of chances across (featured users) has been observed for different initial probabilities
the different RBs. Although this will guarantee the alternation of the PF candidate user. To this end, we keep the total number of
between cell-edge and cell-centric users from one RB to the next users at 420, and the maximum number of FUs at 21. Fig. 2 shows
(the change in chances changes the probability sharply), the algo- the average throughput relative to a pure PF scheduler (on the y-
rithm quickly runs out of chances and mandates a reset for the axis) for different scheduling techniques (on the x-axis) as well
chances as described in EUFS part ðeÞ since the total number of as different initial P(PF) for our proposed EUFS. Probabilities for
chances (i.e. UE_C) is directly proportional to K. Hence, for the total PF candidates less than 70% yield a sharp degradation in cell-
number of RBs, static probability is effectively performed due to centric user throughput since PF users are unlikely to be chosen,
the very frequent reset to the total number of chances. Still with which could not be preferable. Hence, we run the algorithm using
K = 1, as shown in Fig. 1, the dynamic probability assignment initial PF probabilities larger than 70% only. The average through-
through EUFS achieves around 20% performance improvement put for the EUFS algorithm decreases in a sub-linear rate by
over static probabilities. For small K values, the system memory increasing the probability of PF candidate. This is because the num-
is shallow, which implies a behavior closer to fixed probabilities. ber of assigned RBs to cell-edge users decreases as the probability
While increasing the value of K, deeper memory is preserved and of PF candidates is increased. The highest throughput increase
more adaptive RB assignment to the users is performed. Accord- using our algorithm compared to the pure PF algorithm occurs at
ingly, a better performance is achieved. On the other hand, the the lowest probabilities assigned to PF candidates. As seen in
extreme case of increasing the K value as compared to the number Fig. 2, we got an increase of 2x of cell-edge user performance when
of RBs (i.e., K  number of RBs) indicates too much initial chances the initial probability of PF candidates is 70%. The increase is about
for any user (edge-user or cell-centric user). This again reduces the 1.5x when increasing the PF probability between 80% and 90%.
effectiveness of the dynamic probability algorithm (EUFS), since for Next, we examine the average throughput across all cell-centric
each RB only one chance is deducted from the total chances. With a users while decreasing the PF candidates’ probability. As shown in
very high number of total chances, the effect of the dynamic prob- Fig. 2, the average throughput for cell-centric users relative to a
ability changes across RBs is minor and the algorithm again is get- pure PF scheduler (on the y-axis) is decreasing in a sub-linear man-
ting closer to the static probability assignment. Fig. 1 indicates that ner while the probability assigned to the PF candidate (on the x-
the maximum throughput is achieved at K = 140. Accordingly, we axis) decreases. By decreasing the probability of PF candidates,
decide to only consider this value of K for EUFS algorithm in the more RBs are given to cell-edge users and less are given to cell-
following experiments in order not to deviate from the main objec- centric users as compared to a pure PF scheduler. The proposed
tives of these experiments. That is, to compare the proposed algo- EUFS algorithm guarantees that the amount of RBs drawn from
rithm with the state-of-the-art algorithms while keeping the conventional PF candidates, as supposed to be carried out in pure
scaling factor fixed. PF scheduler, to be given to cell-edge users do not degrade the
overall performance of the latter that much because the taken
5.2. Assessing the throughput RBs is drawn in a distributive manner to avoid concentrating the
losses in one or few users which could cause degradation to their
According to 3GPP [6], the user throughput is defined as the accomplished throughputs. It is also shown in Fig. 2 that the largest
ratio of the number of information bits that the user successfully decrease to all cell-centric users throughput occurs at the lowest
receives divided by the total simulation time. If user k has p down- probability for PF candidates. Regarding the EUFS algorithm, the
link packet transmissions during the simulation period ‘‘T(sim)”, maximum decrease in cell-centric user throughput was approxi-
and if there are q(k,i) packets in the ith transmission, and if b(k,i, mately 25%, and this occurred when the initial probability of PF
j) denotes the number of correctly received bits in the jth packet; can-didates was set to 82%. This decrease is shrinking to about
then the average user throughput for user k is: 10% when increasing the probability of PF can-didates beyond 70%.
By investigating the performance of both cell-edge and cell-
X
pðkÞ X
qðkÞ
RðKÞ ¼ ð bðk; i:jÞÞ=TðsimÞ centric users simultaneously while changing the probability of PF
i¼1 j¼1 candidates, the best tradeoff is to assign probabilities to PF candi-

491
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

Fig. 2. Performance comparison for cell-edge/cell-centric users using different scheduling algorithms and fixed number of users.

dates from 80% to 90%. This increases the average throughput for number of UEs/cell, the amount of increase in the average through-
cell-edge users to about 150%, while reducing the average through- put is higher due to the abundance of RBs/users ratio unlike at
put for cell-centric users by only about 10%. Operators could decide higher number of UEs/cell. For instance, at 10 UEs/cell, the average
to further improve the cell-edge performance on the cost of addi- throughput for cell-edge users using EUFS algorithm is about 150%
tional degradation for the cell-centric user performance. To achieve compared to PF algorithm. While at 60 UEs/cell, the average
that, operators could use higher probabilities for PF candidates as throughput for cell-edge users using EUFS algorithm is about only
demonstrated in the experiments. 125% compared to PF algorithm. As explained before, EUFS algo-
Fig. 2 indicates also that our proposed algorithm clearly outper- rithm guarantees that more RBs are given to cell-edge users com-
forms the weighted SNR algorithm from [15] in all cases of used PF pared to the pure PF scheduler, which add to their achieved
probabilities. Our algorithm protects the users from being affected throughput.
sharply by re-adjusting the probabilities when moving from any RB Fig. 3 indicates also that our proposed algorithm is better than the
to the next as discussed in Section 3. However, in the weighted SNR weighted SNR algorithm [15], since EUFS algorithm implies dynamic
algorithm, using fixed weights across the whole TTI does not main- probabilities by re-adjusting the probabilities when moving from
tain the adaptability needed, causing aggressive distribution for any RB to the next as discussed in Section 3. However, in the
the RBs among the users. The same behavior is repeated when weighted SNR algorithm, fixed weights are used across the whole
observing the average throughput of all cell-centric users as shown TTI which does not maintain the adaptability needed, causing
in Fig. 2. aggressive distribution for the RBs among the users. Furthermore,
Furthermore, Fig. 2 indicates that our proposed EUFS algorithm Fig. 3 indicates that our proposed EUFS algorithm outperforms the
outperforms the dynamic PF algorithm from [16] in all cases of dynamic PF algorithm from [16] as updating scheduling priorities
used PF probabilities. Updating scheduling priorities for users to for users to be based on user’s averaged SINR - as performed in
be based on user’s averaged SINR - as performed in dynamic PF - dynamic PF - seems to be less efficient than using user instantaneous
seems to be less efficient than using user instantaneous through- throughput - as used in our proposed algorithm.
put - as used in our proposed algorithm.
Finally, the average throughput of all considered cell-edge users 5.3. Assessing the SNR
(featured users) has been observed for different number of users
per cell. To this end, we keep the initial probabilities of the PF can- In the third experiment, we examine the average SNR for cell-
didate user at 80%. Fig. 3 shows the average throughput for cell- edge and cell-centric users while varying the total number of users
edge users relative to a pure PF scheduler across the different per cell and keeping the number of total RBs fixed at 100 RBs/cell.
scheduling techniques (on the y-axis) versus the total number of The goal is to test the scalability of our proposal compared to other
users per cell as well as the number of cell-edge users per cell algorithms. Figs. 4 and 5 depict the average SNR relative to a pure
(on the x-axis). Our proposed EUFS algorithm outperforms other PF scheduler operating at 60 users/cell (on the y-axis) versus the
scheduling algorithms in all values of number of UEs/cell. When used number of users per cell (on the x-axis) for cell-edge users
comparing EUFS to PF algorithm, we could notice that at lower and cell-centric users, respectively. The average SNR for cell-edge
492
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

Fig. 3. Performance comparison for cell-edge users using different scheduling algorithms and variable number of users.

Fig. 4. Comparing SNR for cell-edge users using different scheduling algorithms.

Fig. 5. Comparing SNR for cell-centric users using different scheduling algorithms.

users in EUFS algorithm is the highest compared to other algo- At lower numbers of users (e.g., 10 users/cell), the increase in
rithms as shown in Fig. 4. In other words, EUFS algorithm is the the average SNR for cell-edge users using the EUFS algorithm is
most beneficial to cell-edge users. This comes at the expense of noticeable, as the average number of RBs/users is high ( 10 RBs/
the SNR for cell-centric users which is the lowest using EUFS algo- user/cell). On the contrary, the percentage of increase in the aver-
rithm, as shown in Fig. 5. However, for the EUFS algorithm, cell- age SNR for cell-edge users using the EUFS algorithm shrinks at
edge users gain 80% increase in the SNR relative to PF at 60 UEs/- higher number of users per cell as the average number of RBs/users
cell, compared to only 45% reduction in the SNR for cell-centric becomes low. By increasing the number of users per cell, the boost
users relative to PF at 60 UEs/cell. of SNR for cell-edge users increases from 55% to 80%. While the loss
493
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

Fig. 6. Fairness Index Comparison.

in SNR for cell-centric users ranges from 45% to 35% compared to ing the PF algorithm as a preliminary step as discussed in Section 3,
PF. Several techniques are available in the literature to enhance it consumes more time than PF algorithm. However, this increase
the achieved SNR more such as COMP [17–20] and heterogeneous do not exceed 7.5% from total PF execution time.
networks [2], which are beyond the scope of this paper.
Fig. 4 indicates also that our proposed algorithm is better than
6. Conclusion
the Dynamic PF algorithm [15] for cell-edge users as the latter
implies a threshold for the maximum time in which the cell-edge
The proposed algorithm has proven to achieve better perfor-
users would be favored within. Hence, using the dynamic PF sched-
mance for cell-edge users compared to PF scheduler used in LTE.
uler, cell-edge users staying relatively long in the cell-edge area are
In addition, it has shown that the side effects on cell-centric users’
assigned lower RBs than when using the EUFS algorithm. The EUFS
performance resulting from decreasing the RBs that they were sup-
algorithm serves cell-edge users as long as they experience poor
posed to take in PF technique are limited. A performance compar-
signal strengths and with no limitation in time, as explained in Sec-
ison with the leading scheduling methodology in LTE has been
tion 3. At the lowest number of UEs/cell used (i.e., 10 UEs/cell), the
quantitatively evaluated for featured and non-featured users. Per-
average RBs per user is 10 RBs/user/cell. Hence, the dynamic PF
formed experiments have been shown that the best trade-off is to
scheduler will compensate the cell-edge users, which have been
assign probabilities to PF candidates from 80% to 90% so as to mul-
punished due to their long stay in the cell-edge area, more effi-
tiply the average throughput for beneficial users approximately
ciently by exploiting the relatively high average RBs per user com-
about 150%. Consequently, this will reduce the average throughput
pared to other cases. As a result, the SNR achieved by the dynamic
for non-beneficial users to only about 10%. Our algorithm could
PF scheduler at 10 UEs/cell is significantly higher and approaching
further take advantage of COMP and joint scheduling techniques
the SNR results of the EUFS algorithm compared to other cases of
to increase the overall system performance among all users.
number of UEs/cell.

Declaration of Competing Interest


5.4. Assessing the fairness

The authors declare that they have no known competing finan-


In the fourth experiment, we use Jain’s fairness index [27] to
cial interests or personal relationships that could have appeared
measure the fairness of the system across all algorithms. Jain’s fair-
to influence the work reported in this paper.
ness index is maximum (i.e., unity) when all users achieve the
same throughput. The closer all user throughputs are to each other,
the larger Jain’s fairness index. Fig. 6 indicates that the EUFS algo- Acknowledgement
rithm outperforms the other algorithms in terms of the fairness
index achieved (i.e., 0.897). According to the definition of Jain’s The authors send their acknowledgement to STDF for their
fairness index provided in Section 4, and the nature of operation financial grant of the Cloud Computing Center of Excellence (num-
for the EUFS algorithm, we could notice that while increasing the ber 5220). This research was supported in part by the Distributed
throughput of cell-edge users and decreasing the throughput of and Networked Systems Research Group Operating Grant number
cell-centric users simultaneously, the two throughputs seem to 150410, University of Sharjah.
be approaching their mean (i.e., the fair throughput). Therefore,
this makes EUFS algorithm maintain a longer term of fairness than References
other methods.
[1] W.N.. D.B.. J.L.. I. Kang. Advanced interference management for 5g cellular
networks. IEEE Commun Mag 2014;52(5):52–60. doi:10.1109/
5.5. Assessing the execution time MCOM.2014.6815893.
[2] J.G.A.. S.B.. W.C.. S.V.H.. A.L.. A.C.K.S.. J.C. Zhang, What will 5g be?, IEEE J Sel
In the fifth experiment, we investigate the execution time Areas Commun 2014;32(6):1065–82. doi:10.1109/JSAC.2014.2328098.
[3] Evolved universal terrestrial radio access (e-utra); user equipment (ue) radio
needed for each algorithm to make sure that the EUFS algorithm transmission and reception, version 10.4.0; 2011. https://www.3gpp.org/
fits our real-time application. Since the EUFS includes implement- [accessed: 2019-8-22].

494
W.S. Afifi et al. Ain Shams Engineering Journal 12 (2021) 487–495

[4] Evolved universal terrestrial radio access (e-utra) and evolved universal Ali A. El-Moursy received the Ph.D. in the area of High-
terrestrial radio access network (e-utran); overall description; stage 2; 2012. performance Computer Architecture from University of
https://www.3gpp.org/ [accessed: 2019-8-22]. Rochester, Rochester, NY, USA, in 2005. Dr. El-Moursy
[5] Evolved universal terrestrial radio access (e-utra): Physical layer procedures. has worked for Software Solution Group, Intel Corp., CA,
v11.11.0; 2013. https://www.3gpp.org/ [accessed: 2019-8-22]. USA till early 2007. In 2007, he has joint Electronics
[6] Third generation partnership project. https://www.3gpp.org/ [accessed: 2019- Research Institute, Giza, Egypt. His research interest is
10-22]. in high-performance computer architecture, multi-core
[7] G.P.. L.A.G.. G.B.. F.C.. P. Camarda, Simulating lte cellular systems: an open multi-threaded mirco-architecture, power-aware
source framework. IEEE Trans Veh Technol 2011;60(2):52–60. doi:10.1109/ micro-architecture, simulation and modeling of archi-
TVT.2010.2091660.
tecture performance and power, workload profiling and
[8] P.K.. S.K.. C. Dabas. Comparative analysis of downlink scheduling algorithms
characterization, cell programming, high performance
for a cell affected by interference in lte network. Ann Data Sci 2016;3(2):135–
53. doi: https://doi.org/10.1007/s40745-016-0076-x [Springer-Verlag Berlin computing, parallel computing and Cloud computing.
Heidelberg]. Dr. El-Moursy has also participated with IBM Cairo Technology Development
[9] Kumar M. Fractional order fir differentiator design using particle swarm Center, Egypt, as a visitor research scientist for the period from Feb. 2007 till Jan.
optimization algorithm. Int J Numer Model Electron Netw Dev Fields 2019;32 2010. In Sep. 2010 Dr. El-Moursy has joint ECE Dep. at University of Sharjah,
(2):e2514. doi: https://doi.org/10.1002/jnm.2514. Sharjah, UAE as an Assistant Prof.. In Jan.2017, Dr. El-Moursy has been promoted to
[10] S.S.. A.A.. M.K.. T.K. Rawat, Adaptive infinite impulse response system the Associate Prof. Rank.
identification using teacher learner based optimization algorithm. Appl Intell
2019;49:1785–802. doi: 10.1007/s10489-018-1354-4.
[11] Aggarwal A, Rawat T, Kumar M, Upadhyay D. Design of optimal band-stop fir Mohamed Saad (Senior Member, IEEE) received the Ph.
filter using l1-norm based rcga. Ain Shams Eng J 2018;9(2):277–89. doi: D. degree in electrical and computer engineering from
https://doi.org/10.1016/j.asej.2015.11.022. McMaster University, Hamilton, Canada, in 2004. He is
[12] L.S. P.W. F. Liu. Particle swarm optimization based resource block allocation currently an Associate Professor at the Department of
algorithm for downlink lte systems. In: 18th Asia-Pacific Conference on
Electrical and Computer Engineering, University of
Communications (APCC), Jeju Island; 2012. p. 970–4.
Sharjah, UAE. His research interests include networking,
[13] F.H.T.V.. B.H.P.G.. F.G.C.R.. L.L.L.. M.V.G. Ferreira. Dynamic resource allocation
communications and optimization, with current activity
in lte systems using an algorithm based on particle swarm optimization and
bmwm network traffic modeling. In: Proceedings of IEEE 6th Latin American focused on the optimal design of wireless and wired
Symposium on Circuits & Systems (LASCAS). Montevideo, Uruguay: IEEE; communication networks, and optimal network resource
2015. p. 1–13. management. He has also held research positions with
[14] E.D.. S.P.. J.S.. P. Beming. 3G Evolution: HSPA and LTE for Mobile Broadband. the Department of Electrical and Computer Engineering,
Academic Press; 2008. University of Toronto, Toronto, Canada, and the Advanced
[15] Y.Y.. D.E.. X.H.. M.M. Downlink resource allocation for next generation wireless Optimization Laboratory at the Department of Computing and Software, McMaster
networks with inter-cell interference. IEEE Trans Wireless Commun University, Hamilton, Canada. Dr. Saad was the recipient of the best paper award in
2013;12:1793–83. doi: 10.1109/TWC.2013.030413.120760. the IEEE Symposium on Computers and Communications, Riccione, Italy, June 2010.
[16] X.N.. V.G.. Z.W.. Q. Yongquan. A dynamic pf scheduler to improve the cell edge He was the recipient of the University of Sharjah ‘‘Annual Incentive Award for
performance. In: Proceedings of IEEE 68th Vehicular Technology Conference Distinguished Faculty Members”, for excellence in research, April 2010 (university-
(VTC 2008-Fall). Moscow, Russia: IEEE; 2008. p. 1–5. wide). He received also two best teaching awards by the IEEE Women in Engineering
[17] Nokia lte smart scheduler white paper. https://onestore.nokia.com/asset/
Society, University of Sharjah (in 2007 and 2009). He was also the recipient of a 2005–
200170?download [accessed: 2019-8-22].
2006 Natural Sciences and Engineering Research Council of Canada (NSERC) post-
[18] S.K.. S.D.. H.M. Gupta, Co-operative downlink scheduling for cell edge and
handoff users. In: Proceedings of IEEE Wireless Communications and doctoral fellowship. He is a senior member of the IEEE.
Networking Conference (WCNC): Mobile and Wireless Networks. Paris,
France: IEEE; 2012. p. 1–6.
[19] G.V.. A.T.. J.J.. M. Juntti. Iterative scheduling for cell-edge in multi-cell mu- Salwa M. Nassar got her PhD in 1984 from Electronics
mimo. In: Proceedings of IEEE Global Conference on Signal and Information and Communication Department, Faculty of Engineer-
Processing (GlobalSIP). Texas, USA: IEEE; 2013. p. 1–6. ing, Cairo University in the field of Parallelism in Pro-
[20] M.A.. M. Mehrjoo. Improving the performance of cell edge users with fractional gramming Languages. Her professional experience
coordinated fair scheduling. Inst Eng Technol (IET) Commun 2014;8 started in 1974 by being a research assistant, in the
(13):2315–21. doi:10.1109/MCOM.2014.6815893. Computer & System Department, Electronics Research
[21] S.M.. J. Rodriguez. Green Communication for 4G Wireless Systems. River Institute. Then, she became an instructor and assistant
Publishers; 2013. professor in Computer & System Department, Elec-
[22] A.-M.B.. N.N.. H.H. Downlink scheduling with economic considerations for tronics Research Institute. She taught in the American
future wireless networks. IEEE Trans Veh Technol 2009;58(2):835–824.
University in Cairo AUC from 1987 to 1997. In 1991, she
doi:10.1109/TVT.2008.927039.
became an associate professor in 1991 and a full pro-
[23] Y.-C.W.. T.-Y. Tsai. A pricing-aware resource scheduling framework for lte
fessor in 1996. She was the head of EU information
networks. IEEE/ACM Trans Netw 2017;25(3):1–14. doi: 10.1109/
TNET.2016.2629501. Point InP. Prof. Nassar is an Ex-ERI president and she is now the PI of the cloud
[24] G.P.. L.A.G.. G.B.. R.F.. P. Camarda. Two-level downlink scheduling for real-time center of excellence and the head of the HPCloud group at ERI. Prof. Nassar is a
multimedia services in lte networks. IEEE Trans Multimedia 2011;13(5):1052– member in the IEEE, IEEE Computer Society, a member of the Information Tech-
65. doi:10.1109/TMM.2011.2152381. nology Academia Collaboration (ITAC) Steering Committee, and a juror in the egypt.
[25] Vienna lte-a simulators, https://www.nt.tuwien.ac.at/research/mobile- wsis-award, 2006, organised by the MCIT, Egypt 2006. Prof. Nassar had 50 publi-
communications/vienna-lte-a-simulators [accessed: 2019-8-22]. cations and has experience and research intersets in the following topics: Parallel
[26] 3gpp lte release-8. http://www.3gpp.org/specifications/releases/72-release-8 processing, Parallel Logic languages, Modeling & Simulation of parallel programs,
[accessed: 2019-8-22]. Distributed Systems, Computer Networks, Parallel Applications, Parallel Virtual
[27] A quantitative measure of fairness and discrimination for resource allocation Machine, Grid Computing, Cluster Computing and Cloud Computing. Prof. Nassar
in shared computer systems; 1984. has led a number of European and USA funded projects.

Wael S. Afifi got his B.Sc. in 2008 from the Electronics Hadia M. El-Hennawy received the B.Sc. and M.Sc.
and Communication Department, Faculty of Engineer- degrees from Ain Shams University, Cairo, Egypt, in
ing, Mansoura University, Egypt. He received the M.Sc. 1972 and 1976, respectively, and the Doctorate of
in the area of High-performance Computing in 2013 Engineering (Dr.-Ing.) degree from the Technische
from the Electronics and Communication Department, Universitat Braunschweig, Braunschweig, Germany, in
Faculty of Engineering, Mansoura University, Egypt. He 1982. Since 1992, she has been a Professor of commu-
has worked after graduation as a research assistant in nication engineering with the Electronics and Commu-
the Electronics and Communication Department, Fac- nications Engineering Department, Ain Shams
ulty of Engineering, Mansoura University till 2009. In University. In 2004, she became a Vice-Dean for grad-
2009 till today, he has joined the Electronics Research uate study and research. In 2005, she became the Dean
Institute, Giza, Egypt. His research interest is in high- of the Faculty of Engineering, Ain Shams University. Her
performance computing and cellular communication research interests include microwave devices and sub-
networks (e.g., 5G). Wael is a reviewer for many journals and conferences. systems, as well as filters and antennas for modern radar and wireless communi-
cations applications.

495

You might also like