You are on page 1of 6

RAN Resource Slicing and Sharing with NOMA for

Latency Reduction in Uplink URLLC Networks


Nadia Imtiaz Jaya Md. Farhad Hossain
Electrical and Electronics Engineering Electrical and Electronics Engineering
Bangladesh University of Engineering and Technology Bangladesh University of Engineering and Technology
Dhaka, Bangladesh. Dhaka, Bangladesh.
nadia.imtiaz.jaya@gmail.com mfarhadhossain@gmail.com

TABLE I. LATENCY AND RELIABILITY FOR DIFFERENT USE CASES OF


Abstract— A major feature of 5G, ultra reliable low latency URLLC [1]
communication (URLLC) network aims to enable mission
critical services with stringent latency constraints. On the other
End-to-end latency
hand, non-orthogonal multiple access scheme (NOMA), a prime Scenario
(ms)
Reliability
candidate for 5G multiple access scheme, utilizes the received
power differences to allow several users to share the same Tactile Internet 1 99.999%
resource blocks (RBs), increasing the throughput, while
negatively affecting successive interference cancellation (SIC) Factory Automation 1 99.9999%
decoding latency of certain users. Therefore, this paper proposes
Remote Control 5 99.9999%
two different NOMA-based user clustering techniques for uplink
transmissions aiming to reduce latency for time stringent Intelligent Transport 10 99.999%
services. Users are clustered and RBs are assigned among the
users in a way to achieve the target latency constraints of users.
Changes in latency and throughputs are thoroughly investigated Other works looked into multiplexing URLLC with
for the validation of the proposed techniques. eMBB and mMTC for optimum resource utilization and
latency reduction [8]
Keywords— slicing; NOMA; latency; class; clustering; This paper proposes mechanisms on using RB slicing
tradeoff and RB sharing among different service classes of URLLC
with NOMA for meeting the latency constraints of different
I. INTRODUCTION services. Two different NOMA-based user clustering
Ultra reliable low latency networks (URLLC), one of the algorithms are investigated such that all classes meet their
most groundbreaking features in 5G which aim to support respective delay constraints. Previously, NOMA-based
the mission critical applications. This network guarantees clustering has been extensively explored for connectivity
the working of mission critical services, such as industry maximization [10], for contention free grant access [7], and
automation, tactile internet, augmented reality, self-driving so forth. Radio slicing [10] had been used to guarantee the
cars, and tele-surgery within their required latency and latency of different service class but not incorporation with
reliability constraints [1]. The different services of the NOMA. Another work [11] explores adaptive radio resource
URLLC class have different bounds of latency and slicing with NOMA but targets user throughput.
reliability, as shown in table 1 [1]. Rest of the paper is organized as follows - in Section II,
This paper asserts to reduce the user plane latency of the we explain our system model, including certain constraints
uplink radio access network (RAN); particularly the packet and assumptions. In this section we have also analytically
processing at base station (BS). Besides the resource describe out two resource allocation models and their
scheduling latency, the RAN user-plane uplink latency objectives. Next in Section III, we verified our claims with
sources for 5G includes transmission time interval (TTI) ȴue tb tc ȴeNB
alignment, propagation delay and signal processing at the ta
UE and BS plus delay due to any retransmission, as shown
Packet sent to
rest of the
network

in Fig 1. The proportion that these delay sources form of the


total uplink RAN delay is also depicted [2]. Much work has Total uplink RAN delay ,TD
been done to reduce the delay in each of these components. TD= ta+ȴUE+tb+tc+ȴeNB+tr
[3] Explores the use of queuing systems and scheduling ta - Packet data generation time at UE
policies to reduce the transmission time intervals. Other ȴUE -Packet processing time at UE
authors have focused on optimizing the decoding and tb – TTI alignment time
encoding of signals by incorporating faster processing units tc – Uplink transmission interval
ȴeNB – Decoding of packet at eNB
(GPUs) instead of (CPUs) [5] and incorporating different
tr – Delay due to retransmission
decoding algorithms [4].
Fig. 1. RAN uplink delay sources.

‹,(((
Authorized licensed use limited to: Istanbul Universitesi. Downloaded on March 31,2022 at 11:59:46 UTC from IEEE Xplore. Restrictions apply.
simulations and explore the characteristic of the two models
for each class. Section IV concludes the paper with a brief
discussion on the key findings.
II. SYSTEM MODEL
A. Overview
We consider a wide area urban macro cell setup with
densely distributed users in outdoor scenario. The BS is
located at the center and is surrounded by N UEs as shown
in Fig. 2. UEs are of M different URLLC service classes –
Class I (CI), Class II (CII) and Class III (CIII), …, Class M
(CM) with increasing order of their delay constraints, i.e. ,
Class I being a more delay sensitive application compared to
Class II which in turn is more latency stringent than Class
III, and so on. Uj i = jth user of Class i, where i = 1, 2, 3, ..M,
j = 1, 2 , 3,...,mi, and mi is the total number of users in Class
i, therefore, m1+m2+…+ mr+..+mM =N. Among the N users,
the different classes exist with equal probability; therefore
the total users in all of the classes are approximately equal.
Also, the UEs are modeled to be uniformly distributed in a
circle of radius r km. The channel gain between the Uj i and Fig. 2. Modeling the outdoor cell scenario
BS is modeled by hji, where hji is 10PL/10. Here path loss, PL Before clustering the users with PD-NOMA, the users of
is the urban macro cell non line-of-sight path loss in dB. different classes were processed in separate ways so that
In both the models discussed below, power domain each class is able to meet its respective delay constraints. In
NOMA (PD-NOMA) has been used as the multiple access the first model, the available network resources were sliced
technique to make the most effective use of the available in three segments – each segment is customized to meet the
spectrum. In uplink PD-NOMA, users send data over the demands of a definite class. 5G radio slicing is a feature
same channel resources but are differentiated from each which enables the available resource blocks of a network to
other in their received power.[4] This power difference be customized to support the diverse requirements of
between the users need to be sufficient so that they can be heterogeneous service classes of a single network. The same
decoded at the receiving end using SIC decoding without radio resource blocks can be tailored to create several virtual
enduring significant error. For maintaining certain reliability networks. [12]
in terms of a given fixed bit error rate (BER), clustering of In the second model, the transmit power of the users are
users in the same RB is done such that the difference regulated such that NOMA clusters always has lower index
between their received power is maximized. At the BS, the users with higher received power decoded earlier.
users in a cluster are decoded in decreasing order of their
received power [4]. Users of higher power are decoded with For illustration purpose and analytical simplification,
interference from all the other users with lower power, the three different classes, CI, CII, and CIII have been
higher this interference the higher the BER and lower the considered.
throughput of that user. After decoding, the users signal is B. Mechanism I
subtracted from the cumulative received signal, effectively In this model, the channels are sliced into three
eliminating its interference from the other users’ signals. As segments, each containing a certain number of resource
a result, the user with the lowest received power does not blocks to serve the three classes. The number of RBs for
experience any interference while higher power users each class is customized in such a way that the lower index
experience the most interference. For example if users in a class users, i.e. users with tighter latency constraint, have a
single RB have power Pj, where j = 1, 2,..., N and Pm > Pn smaller ratio of number of users to number of RB compared
for m < n, then the throughput of user with power Pj is given to higher index. This implies that less number of users is
by : made to share the same RB for lower index class users than
௉௝௛௝
Rj =߱ߚlog2 (1+σಿ ) (1) higher index class users. For achieving this, the total number
ೕశభ ௉௥௛௥ ାே௢
of available resource blocks, C, are divided unequally
among the classes, taking into account the number of users
ߚ = Bandwidth of channel
in each class.
߱ ൌ The number of resource blocks ሺସି௜ሻ
Rj = Throughput of User j. RBs in class i, Wi = σೣ š‫׊ ܥ‬i = 1, 2, 3,…M (2)
భ ௥
No =Noise Power
The term σே௝ାଵ ܲ‫ ݎ݄ݎ‬in equation (1) is the interference in
N = Total number of users
receiver power of jth user from other users in the cluster C = Total number of RBs
whose power is lower than Pj.
.

Authorized licensed use limited to: Istanbul Universitesi. Downloaded on March 31,2022 at 11:59:46 UTC from IEEE Xplore. Restrictions apply.
Class I Class II Class III
(12 RBs) (8 RBs) (4RBs) Figure 4 shows the above clustering scheme for 12 RBs
and 30 users where each class has 10 users.
C. Mechanism II
In the second model the users of different classes are
allocated varying degrees of a reference power Pr. for the
uplink transmission – Class I has transmission power Pr ,
Class II has aPr and Class II has bPr, where a and b are
constants such that 0<b<a<1. Consequently, any user of
12 Resource class I, Uj i , has higher transmit power Pj1 , compared to any
Blocks user of Class II, , Uj2 with power Pj2 which in turn has higher
transmit power than class III, Uj3 with power Pj3 . While the
Fig. 3. Resource slicing of uplink bandwidth. received power at the BS among the users of any class are
proportional to their channel gains. The constants a and b
Equation (2) distributes the number of RBs C among the has been chosen such that ahmax< hmin , and bhmax<ahmin; this
makes sure that all the C1 users will have received power
different Class users, such that the number of users per RB greater than CII, which in turn will have greater power than
ሺସି௜ሻ
increases as the class index increases. The factor cf = σೣ in CIII users.
భ ௥
(2) makes sure that the higher the index of the class, i.e. the After allocating power among the users by the above
more relaxed the latency constraint of the class, the lower is scheme, user clustering for all the users of the three classes
the number of RBs it has out of the total available RBs, C. are done in the same manner as each class of Mechanism I.
Figure 3 illustrates this and has been constructed using an Here, unlike Mechanism I, users are clustered without
example number of 24 RBs. Here, according to factor cf, 12 considering its class, but taking into account its received
RBs has been allocated to CI, 8 RBs to CII and 4 RBs to power only. Since the above power allocation scheme
CIII. Uniform distribution has been used to model the user always ensures all lower index users have greater received
classes and so there will not be significant difference in the power than higher index users, every cluster has CI users as
number of users of different classes. highest power users and CII as intermediate and CII as the
least received power user. Therefore in each cluster, the
The users of each class are then arranged in descending
lower index user is always demodulated first, followed by
order of their received power, Pji and numbered accordingly,
CII user and then CIII. CI users then experience a lower SIC
Pxi >, Pyi and x<y. For Class i, the number of cluster is equal
decoding latency compared to CII users which will have a
to the number of RBs which is Wi and the users are
lower decoding latency compared to CIII. All the users are
distributed in each cluster, such that the users maintains
arranged in descending order of their received power. As
received power distinction and experiences minimum
before Pji represents the power of class i and jth user of that
interference from each other. This ensures that the sum
class, such that Pmi > Pni , where m<n and i has the same
throughput in each cluster is maximized and BER is
value on both the sides. The users of the three classes, with
minimized. For example in Class I with W1 RBs and K1
CI having K1 users CII having K2 users and CIII having K3
users, the users clustering will look like:
users, are then arranged within the W clusters which is equal
{Here, the constraint 2W1< K1<3W1, is used for ease in to the number of available RBs in the following manner :
illustration only.} {For illustration purpose the following constraints are
applied - K1=W+R-1, K1+K2=2W+T-1, and
ଵ ଵ
Cluster 1: ܲଵଵ , ܲଵା௪ భ , ܲଵାଶ௪భ K1+K2+K3=3W+S-1, where R<S<T<W.}
ଵ ଵ ଵ
Cluster 2: ܲଶ , ܲଶା௪భ , ܲଶାଶ௪భ
: Cluster 1: ܲଵଵ , ܲௐାଵ

,

ܲௐିோାଶ ଷ
, ܲௐି்ାଶ
ଵ ଵ ଵ ଵ ଶ ଷ
Cluster R: ܲோଵ , ܲோା௪ భ , ܲ௄ భ Cluster 2: ܲଶ , ܲௐାଶ , ܲௐିோାଷ , ܲௐି்ାଷ
: :
Cluster W1: ܲௐ ଵ ଵ
భ , ܲଶௐ భ , 0 Cluster R: ܲோଵ , ܲଵଶ , ଶ
ܲௐାଵ ଷ
, ܲௐି்ାோାଵ
:
Cluster S: ܲௌଵ , ܲௌିோାଵ
ଶ ଶ
, ܲௐାௌିோାଵ , 0
:
Cluster T: ்ܲଵ , ்ܲିோାଵ
ଶ ଷ
, ܲଵ , 0
:
ଵ ଶ ଷ
Cluster W: ܲௐ , ܲௐିோାଵ , ܲௐି்ାଵ , 0.

Figure 5 depicts this clustering for 10 CI users, 10 CII


users, and 10 CIII users, whose respective powers are ܲଵଵ to

ܲଵ଴ , ܲଵଶ to ܲଵ଴

, and ܲଵଷ to ܲଵ଴

.
Fig. 4. Mechanism I clustering with 12 RBs and 30 users.

Authorized licensed use limited to: Istanbul Universitesi. Downloaded on March 31,2022 at 11:59:46 UTC from IEEE Xplore. Restrictions apply.
Clusters (12) reduce the latency for more latency stringent services while
trading it with increased latency for services that have
relaxed delay constraints.
1 2 3 4 5 6 7 8 9 10 11 12 In Mechanism I, since lower index users are allocated
larger number of RBs for the same user number, cluster size
P11 P21 P31 P41 P51 P61 P71 P81 P91 P101 P12 P22 for CI users is lower than CII which in turn is lower than
P32 P42 P52 P62 P72 P82 P92 P102 P13 P23 P33 P43 CIII as shown in Fig.4. Since smaller clusters have in
P53 P63 P73 P83 P93 P103 general smaller average decoding time than larger ones, the
average decoding time for the classes follow the pattern
CI<CII<CIII as simulations show in Fig.6. For Mechanism
II as Fig.5 shows the decoding order follows CI first, then
CII, and then CIII; the lower index class has smaller
CI USERS CII USERS CIII USERS decoding time than higher index class for being higher up in
the decoding order. The results shown graphically in Fig.6
Fig. 5. Mechanism II clustering with 12 RBs and 30 users.
validate the clustering in Mechanism II.
Comparing the graphs for both the mechanisms in Fig.6 ,
III. SIMULATION RESULTS we see that Mechanism II has given better results than
The two models are simulated in the frequency band of Mechanism I in all the three classes. In Mechanism II, the
4GHz using 24 resource blocks of 200 kHz. The cell sector length of the cluster has no effect on the decoding time of
simulated is modeled to be in an urban macro scenario with the users that are decoded initially in the cluster. This is only
a cell radius of 1km following the 5G NR standard for valid when the decoding time at BS is not the limiting factor
smaller and denser cells. The simulations are done for UE in the total uplink delay time, TD as shown in Fig.1. In the
number ranging from 100 to 800. One unit of time is system model considered, tc, uplink transmission time forms
equivalent to the time needed to decode a single user at the the limiting factor in the total uplink delay.
SIC decoding. For Mechanism II, value of a taken was 0.8 B. Throughput against User number
and value for b was 0.3.
Throughputs, given by equation (1), are evaluated against
A. Average decoding time against User number increasing user number for both the mechanisms and are
The average decoding time for the SIC decoding at the shown graphically in Fig.7 and Fig.8. As equation (1)
BS is calculated for each service class in both the shown, throughput depends on the received power of the
mechanisms. The graphs represent the average decoding user and the collective interference from the other users. In
times for each class (CI, CII, &CIII) of both the mechanisms Mechanism I, as Fig.4 shows, lower index user experience
(Mechanism I and Mechanism II). Also included is the lower interference compared to higher index user due to
decoding time of users when they are clustered randomly being clustered into shorter clusters, with less number of
without taking into account their class or model of the user. other users to cause interference, while the received power
The decoding time of the classes increases as the class index plays no role due to it being proportional to channel gains,
increases for both the mechanisms. Also, in both the which is randomized in our model. As Fig.7 depicts, the
mechanisms, the lowest index class has decoding time lower throughput follows the following sequence CI>CII>CIII at
than the results for random clustering, and the highest index any user number.
class, CIII has higher decoding time than the time for
random clustering. Therefore, both the models are able to

Fig. 7. Throughput against user number for Mechanism I users.

Fig. 6. Relation of average decoding time against user number for the different
clustering algorithms.

Authorized licensed use limited to: Istanbul Universitesi. Downloaded on March 31,2022 at 11:59:46 UTC from IEEE Xplore. Restrictions apply.
As a result for CII and CIII, lower decoding times are more
abundant when following Mechanism I than Mechanism II.
The crossover in the time CDF graphs for the two models in
CII and CIII classes demonstrates this skewness. For CI
users, in both the Mechanisms, users have a significant and
comparable portion being decoded initially or early in their
clusters, albeit with Mechanism II having a higher portion of
it than Mechanism I, shown schematically in Fig.4 and
Fig.5. This is reflected in CI - Mechanism II users having a
smaller range for time CDF than CI - Mechanism I, as
already mentioned.
IV. CONCLUSION
In this paper we have suggested two novel resource
allocation models for an urban outdoor scenario with three
different service classes. Heterogeneous service classes
Fig. 8. Throughput against user number for Mechanism II users.
sharing the same network resources are a prime feature of
5G URLLC networks. Our paper aims to develop two
In Mechanism II, as plotted in Fig.8, CI and CII users mechanisms which regulate the latency of three different
have an intertwined throughput graphs while CIII has a service classes with varying latency constraints such that
higher throughput than the other two. As already mentioned, each class is able to maintain its latency boundaries. The
CI users are given higher power than CII which in turn is mechanisms have been explained analytically and the
given higher power than CIII, so that the decoding order network variables, decoding time, its CDF, and throughput
improves latency with respect to each class. From SIC have been simulated over increasing user number; the results
decoding we know that users in each cluster that are of which have also been explained. Decoding time graphs
decoded first have the highest power in each class and depict the decoding time tradeoffs between latency stringent
experience the most interference. As a consequence, CI and and latency relaxed classes, such that each class is decoded
CII users have almost the same throughput, since the ratio of within its latency constraints. Throughputs against user
higher received power to higher interference for CI is number graphs show how the two models affect throughput
approximately close to the ratio of lower receiver power to of the different class. The time CDF graphs also demonstrate
lower interference for CII. CIII users on average have to the range and skewness of decoding times for different
experience very little to no interference, as depicted by the classes in the two models. The simulation results can be
example in Fig.5, and so has the highest throughput used to determine the model suitable for particular network
compared to the other two, shown in Fig.8. The actual requirements of throughput and decoding time limits.
position of the graphs is dependent on the value of a and b
Mechanism I uses resource slicing to dedicate a certain
and hmax.
number of RBs to the different classes while Mechnism II
C. Time CDF allows all the RBs to be dedicated to all the service class
Time cumulative density functions (CDF) for the three while arranging the cluster in order of higher priority to
lower.
different classes in both the mechanisms are simulated and is
graphically shown in Fig.9. The maximum decoding time or
range of time CDF for Mechanism II is always lower than
Mechanism I for each of the classes. This is because for
each class total RB available is lower for Mechanism I than
Mechanism II and so average decoding order is greater for
Mechanism I in each of the classes, evidenced in Fig.4 and
Fig.5; with greater average decoding order in Mechanism II,
maximum decoding time is greater. This result is also
reflected in the average decoding time against user number
for the two models in fig.6, where Mechanism I always has
greater latency than Mechanism II.
Also as shown in Fig.9, the time CDF for CII and CIII
classes is more positively skewed for Mechanism I than
Mechanism II. Mechanism I users of CI and CII have a
higher portion of users with low decoding time compare to
corresponding user of Mechanism II. Fig. 4 and Fig.5
explains this with an example of 12 RBs. Here, almost all
CII and CIII users in Mechanism II are decoded second or
later compared to Mechanism I where a significant portion Fig. 9. Comparison of CDF for the different classes and models.
of CI and CII users are decoded first in SIC decoding at BS.

Authorized licensed use limited to: Istanbul Universitesi. Downloaded on March 31,2022 at 11:59:46 UTC from IEEE Xplore. Restrictions apply.
REFERENCES

[1] Zexian Li, Hamidreza Shariatmadari, Bikramjit Singh, Mikko A.


Uusitalo,”5G URLLC: Design Challenges and System Concepts”, 2018
15th International Symposium on Wireless Communication Systems
(ISWCS)
[2] Amitav Mukherjee,”Energy Efficiency and Delay in 5G Ultra-Reliable
Low-Latency Communications System Architectures”, IEEE Network
vol. 32 no. 2, pp 55-61, March 2018
[3] Apostolos Destounis, Georgios S. Paschos, Jesus Arnau, Marios
Kountouris, “Scheduling URLLC Users with Reliable Latency
Guarantees,” 2018 16th International Symposium on Modeling and
Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt), May
2018
[4] Z. Ding et al., ‘‘A survey on non-orthogonal multiple access for 5G
networks: Research challenges and future trends,’’ IEEE J. Sel. Areas
Communication., vol. 35, no. 10, pp. 2181---2195, Oct. 2017.
[5] Talgat Manglayev, Refik Caglar Kizirilmak, Yau Hee Kho,Nor Asilah
Wati Abdul Hamid, “GPU Accelerated Successive Interference
Cancellation”, Springer Nature 2018
[6] Xiangming Meng ; Yiqun Wu ; Chao Wang ; Yan Chen, “A Universal
Receiver for Uplink NOMA Systems,” 2018 IEEE/CIC International
Conference on Communications in China (ICCC Workshops), August
2018
[7] Seda Doga, Armed Tusha, Huseyin Arslan, “NOMA with Index
Modulation for Uplink URLLC through Grant-Free Access,” IEEE
Journal of Selected Topics in Signal Processing ( Early Access ) May
2019
[8] Guillermo Pocovi, Klaus I. Pedersen,and Preben Mogensen, “Joint Link
Adaptation and Scheduling for 5G Ultra-Reliable Low-Latency
Communications IEEE Access ( Volume: 6 ) pp - 28912 – 28922, May
2018.
[9] Ahmed Elhamy Mostafa, Yong Zhou, and Vincent W.S.
Wing,”Connectivity Maximization for Narrownamd IoT Systems with
NOMA”.
[10] “Guolin Sun ; Guohui Wang ; Guisong Liu, “Air-Interface Slice Based
Dynamic Resource Reservation for Ultra-Low-Latency IoT
Transmissions”, 2016 IEEE 41st Conference on Local Computer
Networks (LCN), Nov 2016
[11] L. Tang, Q. Tan, Y. Shi, C. Wang and Q. Chen, "Adaptive Virtual
Resource Allocation in 5G Network Slicing Using Constrained Markov
Decision Process," in IEEE Access, vol. 6, pp. 61184-61195, 2018.
[12] Adlen Ksentini, Pantelis A. Frangoudis, Amogh PC, Navid Nikaein,
“Providing Low Latency Guarantees for Slicing-Ready 5G Systems via
Two-Level MAC Scheduling” vol 32 no. 6, pp-116-123, December 2018

Authorized licensed use limited to: Istanbul Universitesi. Downloaded on March 31,2022 at 11:59:46 UTC from IEEE Xplore. Restrictions apply.

You might also like