This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE SYSTEMS JOURNAL 1

Fair-QoS Broker Algorithm for Overload-State

Downlink Resource Scheduling in LTE Networks
Nasim Ferdosian, Member, IEEE, Mohamed Othman, Member, IEEE,
Borhanuddin Mohd Ali, Member, IEEE, and Kweh Yeah Lun

Abstract—Efficiently scheduling various service classes while Generally, the tens of megabits of channel bandwidth
fulfilling performance targets is one of the main challenges in 3GPP provided by the LTE networks are still limited by physical
Long-Term Evolution (LTE) communication systems. However, laws [4] and may not be fairly shared among different services
Quality of Service (QoS) and fairness provisioning for all different
service classes are generally unstable due to insufficient radio [5]. This is because greedy applications such as Voice over IP
resources, especially in heavily loaded states of the network, which (VoIP), live video transmission, and download accelerators may
leads to the deterioration of overall system performance. In this overwhelm the network, consume a large amount of resources,
paper, we address these challenging issues by proposing a two-level and consequently prevent other services from getting them.
downlink scheduling algorithm to deliver all traffic types while Therefore, the overload states that occur often in the network
attempting to provide a compromise for the LTE performance
targets. For the higher level of the algorithm, a coherent resource contribute to the unfair access problem of different service
distribution approach is developed by applying a game theory bearers and degrade QoS [6]. These service quality degradation
model in a nested manner to provide per-class fairness. At the and unfair resource usage issues pose a tremendous challenge
lower level, the greedy-knapsack algorithm is properly tailored to for resource-scheduling algorithm designers. A wide range of
optimally allocate resources to the best potential bearers for QoS resource-scheduling issues in LTE networks has been studied
and throughput enhancement. The performance of the proposed
algorithm was evaluated for normal and overload states of the extensively [7], [8]; however, most of these solutions are unfair
network. The simulation results clearly demonstrate that the in essence or ensure only a particular level of QoS requirements.
proposed scheduling algorithm, compared with the reference Moreover, typical resource-scheduling algorithms consider a
scheduling algorithms, provides the best tradeoff for fairness, special kind of traffic, and most of them focus on real-time
throughput, and QoS performance in terms of packet loss rate services. For example, the algorithms proposed in [9] and [10]
and packet latency for different service classes.
mainly focus on real-time multimedia traffic while discarding
Index Terms—Fairness, greedy-knapsack algorithm, LTE down- other kinds of services such as Web browsing or HTTP.
link scheduling, nested game theory, QoS. Accordingly, the objective of this study is to propose an
efficient resource-scheduling algorithm that simultaneously
I. INTRODUCTION
supports fair channel sharing, per-class QoS, and throughput
HE Long-Term Evolution (LTE) cellular communication
T system has emerged as a fast-growing prevalent technol-
ogy delivering a diversity of mobile broadband services in the
enhancement without having to modify LTE specifications. The
resource-scheduling procedure of the time-domain scheduling
algorithm is decoupled from that of the frequency-domain
communication world [1]. As the LTE communication networks scheduling algorithm. Hence, the focus of this paper is the phase
face increasing public deployment along with the growing de- at which the time-domain packet scheduling is involved. The
mand for a wide variety of applications, it is important to support proposed scheduling algorithm, called Fair-QoS Broker (FQB),
their diverse requirements [2] and ensure that resources are fairly contains two main components: a fair resource distributor
distributed among them. However, the LTE resource-scheduling (FRD) and QoS Level Allocator (QLA).
strategy, which is an effective strategy for fulfilling the Quality The FRD is intended to guarantee a FRD among the QoS
of Service (QoS), fairness, and throughput targets, has been left classes. Toward this aim, the FRD component develops a sys-
to the equipment vendors as an open issue [3]. tematically and theoretically coherent decision-making process
by applying a game theory model in a nested manner to fairly
define an answering portion of resources for each class of bear-
Manuscript received December 21, 2016; revised March 10, 2017; ac-
cepted May 3, 2017. This work was supported by the Malaysian Ministry ers. In contrast, QLA monitors the QoS features of each bearer
of Education under the Fundamental Research Grant Scheme FRGS/2/2014/ and accordingly allocates the resource portion predefined for
ICT03/UPM/02/3. (Corresponding author: Nasim Ferdosian.) each class to all bearers belonging to that class. Within each
N. Ferdosian, M. Othman, and K. Y. Lun are with the Department of
Communication and Network Technology, Universiti Putra Malaysia, Serdang bearer class, the QLA selects the best potential bearers to which
43400, Malaysia (e-mail: [email protected]; [email protected]; resources should be allocated using a ranking function that
[email protected]). aggregates the influential QoS parameters and their threshold
B. M. Ali is with the Department of Computer and Communication Sys-
tems, Universiti Putra Malaysia, Serdang 43400, Malaysia (e-mail: borhan@ values.
upm.edu.my). The QoS enhancement does not undermine per-class fairness
Digital Object Identifier 10.1109/JSYST.2017.2702109 because the FRD apportions the channel resources to the bearer

1937-9234 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE SYSTEMS JOURNAL

classes by considering their optimal interactive effects. Further-

more, the FRD does not decrease resource utilization because
it accommodates bearers based on the total resource usage and
bearers’ resource demands. The combination of FRD and QLA
is shown to be effective and convergent, offering a reasonable
tradeoff among fairness, QoS, and throughput enhancement for
different classes of services.
In addition to the above contributions, we derive a model
for evaluating fairness performance. Here, we deal with per-
class fairness because we plainly believe that this concept is
the most meaningful and the closest to LTE QoS expectations.
The need for this model is mainly motivated by the fact that
existing fairness analytical models do not consider the effect of
QoS parameters. Fig. 1. Typical network topology.
The structure of this paper is organized as follows. In
Section II, we overview the optimization-theoretical ap- then proposed a heuristic solution through the Nash Bargain-
proaches for LTE resource-scheduling in the literature. ing Solution (NBS), aiming to provide proportional fairness. In
Section III describes the system model used in our study and a cooperative NBS approach, players negotiate with the base
gives a formal description of the problem. Section IV introduces station and one another to decide the scheduling. However, the
the FQB resource allocation algorithm with the details of its bargaining game among players imposes an additional overhead
components. Section V presents the simulation environment of information transfer needed to reach an efficient agreement
and parameters. In Section VI, a method to evaluate the on scheduling decisions.
per-class fairness level is presented and the performance results Brehm and Prakash [18] formulated the LTE downlink
are discussed. Finally, Section VII concludes this paper. scheduling problem as a fractional knapsack optimization prob-
lem. In this regard, Ferdosian et al. [19] presented the greedy-
knapsack algorithm, which provides an optimal solution to the
II. OPTIMIZATION-THEORETIC SCHEDULING APPROACHES
fractional knapsack problem of downlink resource allocation to
The LTE resource scheduling problem is defined as a problem assure the QoS requirements of bearers from different classes
that cannot be solved optimally to any promising bound by a of service. However, considering QoS requirements without any
deterministic approach without complexity [11]. Heuristic and concern for the extent of their required resources and fair bearer
meta-heuristic approaches have been recognized as efficient so- resource sharing does not lead to an efficient scheduling solution
lutions for these computationally hard problems. We point out for total system performance.
some of the main meta-heuristic methods that have been used Therefore, the lack of a unique solution to treat all three sys-
for the LTE scheduling optimization problem. tem performance targets together as a single scheduling problem
Aydin et al. [12] tried to solve the multiuser scheduling prob- and create a tradeoff among them without having to modify LTE
lem using simulated annealing and genetic algorithms, which specification forms a major challenge in LTE networks. There-
are combinatorial optimization approaches. They provided near- fore, an efficient resource scheduling algorithm must distribute
optimal solutions in terms of average total bit rate without con- the available resources among the active bearers in such a way
cern for the QoS requirements. To capture the QoS requirements, as to provide fair service coverage, maximize throughput, and
a suboptimal algorithm was proposed in [13] based on the binary fulfill QoS requirements over the whole system.
integer programming problem for three types of traffic. Among
QoS-aware schedulers in heterogeneous traffic networks, this III. BACKGROUND AND PROBLEM FORMULATION
algorithm considers both standard specific and individual user
A. Downlink System Model
QoS constraints. In [14], the uplink scheduling problem along
with the consecutive-block constraint was expressed as an in- In the present paper, we consider the downlink scenario of
teger programming problem. A heuristic localized gradient ap- a single cellular network, which is composed of an E-UTRAN
proach was introduced in response to this problem. It showed NodeB (eNodeB, the base station in LTE networks) and units of
that the heuristic version of the localized gradient algorithm cellular user equipment, as depicted in Fig. 1.
outperforms the nonheuristic one with improved normalized The downlink scheduling algorithm relies on the QoS class
throughput. based architecture at bearer level, where a bearer is ascertained
Game theory presents an effective analytical tool to address to convey a certain data flow. Each bearer has its own service and
Pareto-optimal problems [15]. In this regard, Shams et al. [16] performance characteristics depending on the application that
modeled the Orthogonal Frequency-Division Multiple Access generated it. The 3GPP specification [3] standardized them into
(OFDMA) subcarrier scheduling by adopting the coalitional- nine classes (as given in Table I) based on their QoS features to
based game theory to address the fairness issue, compromised ensure that the individual bearers that belong to the same class
by the target user’s data rate. Yaacoub and Dawy [17] formu- of services receive the same minimum level of service quality.
lated the LTE uplink scheduling as a game theory problem and Therefore, a common rate and scheduling policy must be de-
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERDOSIAN et al.: FAIR-QoS BROKER ALGORITHM FOR OVERLOAD-STATE DOWNLINK RESOURCE SCHEDULING IN LTE NETWORKS 3

TABLE I
STANDARDIZED QoS CHARACTERISTICS IN QCI CLASSES [24]

QCI Bearer Type Priority Packet Delay Budget (ms) Packet Error Loss Rate Example Services

1 GBR 2 100 10 Conversational voice

2 4 150 10 Conversational video (live streaming)
3 3 50 10 Real-time gaming
4 5 300 10 Nonconversational video (buffered streaming)
5 Non-GBR 1 100 10 IMS signalling
6 6 300 10 Video (buffered streaming) TCP-based (e.g., www, e-mail, chat, ftp, p2p file sharing,
progressive video, etc.)
7 7 100 10 Voice, video (live streaming) interactive gaming
8 8 300 10 Video (buffered streaming) TCP-based (e.g., www, e-mail, chat, ftp, p2p file sharing,
progressive video, etc.)
9 9 300 10

voted to all bearers assigned to an especial QoS Class Indicator In terms of system performance, the most common optimiza-
(QCI). Accordingly, we consider the scheduling algorithm that tion objective for LTE systems is to distribute resource blocks
fulfills the predefined QoS constraints in the LTE Medium Ac- among different users in such a way that the overall data rate of
cess Control (MAC) layer and thoroughly covers all QoS classes the system is maximized, as stated in (1), where RBi ⊂ RB is
except QCI class 5, which has the highest priority (i.e., one) and the set of resource blocks dedicated to bearer i for transmitting
is especially given the Internet protocol Multimedia Subsystem data, ri,r b (t) is the data rate achieved by bearer i over the rbth
(IMS) signaling mechanism, which is assumed to be completely resource block during a given time interval t, and T denotes the
independent from the access network. number of time intervals.
In this context, the 3GPP-LTE downlink system is based on An optimum level for the total system performance, without
an OFDMA access strategy associated with the eNodeB, while any concern for fairness, might prevent all users in the network
the OFDMA radio resources of the LTE system are expanded from having a satisfactory level of network performance. One
over a time–frequency grid. Let N = {1, 2, . . . , n} be set of the fairness paradigm is to grant each user a long-term and a cer-
bearers waiting to be scheduled by the set of available system tain amount of the system throughput, as stated in (2), where
resource blocks RB = {1, 2, . . . , z} in the tth transmission time ri (t) denotes the overall data rate of bearer i at time t. This
interval (TTI). is computed by summing the bearer’s data rate over the whole
set of resource
 blocks assigned to that bearer at the given time
B. Problem Formulation (ri (t) = r b∈RBi ri,r b (t)). Here, ϕi is the minimum fraction of
The aim of the eNodeB scheduling algorithm is to allocate the total system throughput r̄, needed by bearer i, while ϕi ≥ 0
resources to the bearers in such a way that optimizes the total sys- (ϕi = 0 when bearer i does not have any packet to transmit) and
 n
tem performance. Here, the system performance is determined i=1 ϕi ≤ 1.
in terms of three main LTE targets: throughput maximization, Equation (3) expresses the third pertinent optimization objec-
QoS service guarantee, and fairness provisioning. Inasmuch as tive, which implies that bearers’ average loss and delay should
these targets are in conflict with each other, the scheduling algo- be minimized over multiple time intervals, where di (t) and li (t)
rithm must provide a compromise among them. Therefore, the are the measured packet loss and delay, respectively, of bearer i
scheduling problem for allocating resource blocks to the bearers over time interval t.
in the context of LTE systems can be mathematically formulated Inasmuch as the bearers from different QoS classes have
as a multiobjective optimization problem as follows: distinct predefined QoS specifics, the scheduling algorithm must
fulfill constraints (4) and (5) in order to meet class-based QoS
 
∀t, 1 ≤ t ≤ T : max ri,r b (t) (1) requirements. They imply that packet error loss rate li k and
i∈N r b∈RBi packet delay budget di k experienced by bearer i from QCI class
k (k ∈ K = {1, 2, . . . , 9}) should be less than predefined class-
1
T
∀i ∈ N : lim inf ri (t) ≥ ϕi r̄ (2) based loss and delay threshold, Lk and Dk , respectively. The
T →∞ T t=1 constraint in (6) states that the assigned RB sets of individual
bearers are disjoint sets, thus the resource blocks are assigned to
1 1
T T
∀i ∈ N : min li (t) and min di (t) (3) each bearer exclusively and a single resource block is not shared
T t=1 T t=1 by different bearers during the same time slot.
subject to
IV. FQB ALGORITHM
∀i ∈ N, k ∈ K : di k < Dk (4)
As described in Section III-B, LTE scheduling is multiob-
∀i ∈ N, k ∈ K : li k < Lk (5)
jective in its inclusive formulation, where any improvement
∀i, j ∈ N, i = j : RBi ∩ RBj = ∅. (6) with respect to one objective causes an impairment with respect
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE SYSTEMS JOURNAL

Fig. 2. Schematic diagram of the FQB.

to another one. Typically, multiobjective combinatorial opti- A. Fair Resource Distributor (FRD)
mization approaches are addressed using the Pareto optimiza-
The design goal of the FRD component as a class-based
tion theory. However, in LTE scheduling, a quickly computed
fairness control mechanism is to satisfy the following issues.
scheduling plan that provides a tradeoff among the objectives
1) The allocation of adequate bandwidth for all QCI classes
is preferable to detecting many Pareto points in the problem
so that the resource quota destined for each class is regu-
area and dealing with all the conflicting objectives at the same
lated tightly around the transmission requirements of that
time. Therefore, we propose a multilevel scheduling algorithm
class.
called FQB while considering a single objective at each separate
2) Fair resource allocation, even in overload states of the
level of the scheduling procedure, which is efficient for prac-
network, when the bearers’ demands are bigger than the
tical implementation. FQB, which is composed of two main
available resources.
components, the FRD and QLA (see Fig. 2), aims to guar-
3) The utilization of the available resources should be as
antee the QoS requirements of bearers from different service
efficient as possible. Regulating the QCI class resource
classes and provides fairness without decreasing the total system
quota should prevent a service bearer from using more
throughput.
than its fair share of resources.
As can be seen in Fig. 2, bearers from heterogeneous traffic,
4) The bearers marked as GBR should be served with the
buffered at the eNodeB, enter the FRD component. In the first highest priority so that their predefined data rate require-
step, the buffered heterogeneous traffic is categorized into guar-
ment is satisfied.
anteed bit rate (GBR) and non-guaranteed bit rate (non-GBR)
1) Game Theoretic Approach for FRD: The FRD resource
categories and then classified into the predefined QCI classes, distribution mechanism is modeled by applying game theory in
which are presented in Table I. Thereupon, the free system re-
a layered manner, called here a nested game model. As can be
source blocks that can be scheduled for transmitting the bearers
seen in Fig. 3, the resource competitions among the QCI classes
are computed. In the next step, the FRD component partitions in the lower layer are sequential subgames that are nested in
the free system resource blocks among the different classes of
a larger game between the GBR and non-GBR categories in
bearers to provide per-class fairness using the nested game the-
the upper layer. This section describes the games by defining
ory. The calculated per-class resource quotas are then sent to
the players, competition resources, and outcome of each game.
the QLA component. The QLA monitors the QoS features of
Subsequently, the players’ interactions and convergence of the
each bearer and allocates the defined resource portion of each
games are discussed in details in the next section.
class to the bearers belonging to that class. In each service class,
Let a game be defined by a pair (G, n), where G is the com-
the QLA component selects the best potential bearers to which
petition resource and n is the number of players. As shown
resources should be allocated based on the divergence between
in Fig. 3, the nesting game (B, 2) is a two-player competition
the bearers’ measured QoS parameters and their corresponding
between GBR and non-GBR categories of QCI classes as the
threshold values.
game’s players, with the player set P = {GBR, NonGBR} and
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERDOSIAN et al.: FAIR-QoS BROKER ALGORITHM FOR OVERLOAD-STATE DOWNLINK RESOURCE SCHEDULING IN LTE NETWORKS 5

Fig. 3. Framework of the nested game model.

claim vector C = {cGBR , cnon-GBR }. Each player’s claim is sum It is important to note that each applied game is a one-shot
of the resource requirements of the QCI classes inside the cor- game in which no player-specific information is revealed to
responding category. The outcome of this game is the resource the opponents. Unlike other widely employed game-theoretic
quotas φGBR and φNonGBR granted to the GBR and non-GBR cate- concepts (e.g., the Nash equilibrium) that may require a number
gories, respectively. of iterations before converging to an acceptable solution, the
Subsequently, in the lower layer of the nested game model, equilibrium point of our games can be always achieved in one
there are two four-player subgames: nested-game1 (φGBR , 4), shot. This game is played once and ends with the Shapley value
with player set P = {QCI1, QCI2, QCI3, QCI4} and claim vec- solution strategy vector of all players, when the actions of all
tor C = {cQCI1 , cQCI2 , cQCI3 , cQCI4 } and nested-game2 (φNonGBR , 4), players have been chosen and payoffs received.
with player set P = {QCI6, QCI7, QCI8, QCI9} and claim vec- 2) Nesting and Nested Game Implementation: Considering
tor C = {cQCI6 , cQCI7 , cQCI8 , cQCI9 }. These games are competitions Fig. 3, we describe the nesting and nested games’ implemen-
among the QCI classes to distribute the resource quota of each tation and the computation of their corresponding outputs. We
category among themselves based on their collected bearers’ re- characterize both games using the cooperative coalitional game
source demands. The resource requirements of the bearers can theory to represent the competitive and collaborative scenarios
be quantified in terms of their outgoing packet buffer’s size. among the resource demanding entities, which are the players
The outcome of these games, which is the outcome of the FRD in the games. Players negotiate and cooperate with each other to
component, is a vector of the data rate quotas granted to each make a decision. The cooperating groups of players are referred
individual QCI class. to as coalitions and denoted as S ⊂ P , where P = {1, 2, . . . , p}
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE SYSTEMS JOURNAL

is the set of all players and p = |P |. Every coalition has a worth Eventually, each game converges to the equilibrium point by
in the game, which is quantified by the coalition value v. In the calculating Shapley value φi for each player i inside that game.
characteristic form of coalitional game, value v is defined us- In this way, the FRD component outputs a class-based resource
ing a characteristic function that assigns a real number to every distribution vector φ = {φ1 , φ2 , . . . , φ9 }, where, for example,
coalition S. φ1 is the quota of the resource blocks granted to QCI class 1.
Because in this paper, we focus on heavy-traffic scenarios Given this achievable quota of resources for each QCI class,
where the number of available resource blocks is less than the then the resource allocation for bearers’ transmissions inside
sum of the user resource claims, we map the resource distribu- each class is determined by the QLA component.
tion problem and subsequently the corresponding nesting and
nested games to the bankruptcy model of game theory. The prob- B. QoS-Level Allocator (QLA)
lem is how to obtain an equilibrium point for which the available
amount of resource blocks B ∈ IN is divided fairly among the The design strategy of the QLA component follows this main
set P of players. Let ci ∈ IN denotes player i’s resource demand goal: A weighted allocation of the predefined resource quota for
each QCI class to the bearers of that class with an adjustment of
 c ≡ (ci )i∈P be the nonnegative vector of demands, such that
and
each bearers service levels to guarantee its associated data rate
i∈P ci ≥ B. The characteristic function of bankruptcy game
E that we are considering here is then defined as [20] and QoS requirements. Hence, the QLA responds to these issues
⎛ ⎞ through the following steps (pseudocode of QLA is presented
 in Algorithm 1).
vE (S) = max ⎝0, B − cj ⎠ (7) 1) Compute the current values of the QoS parameters (loss,
j∈
/S delay, and QoS priority), queue length, and data transmis-
sion requests for each bearer at the beginning of every
for every possible coalition S ∈ 2 , where vE (P ) = B and
p
transmission time.
vE (S) are the amounts of resources that are left if all players
2) Assign sufficient resource blocks to the GBR bearers to
outside coalition S receive their claim.
guarantee at least their predefined required data rate value
After defining the characteristic function, the next step in the
and at most their maximum bit rate value. (Bearers uti-
game procedure is to define a content solution to distribute the
lizing the GBR services must not experience throughput
total available resources B among the players using an appro-
degradation caused by overload states.)
priate fairness rule. The Shapley value approach is one of the
3) Perform an overall evaluation for every bearer to rank
most popular solutions useful for games with the aim of fair
them by using a multicriteria ranking function.
allocation. The Shapley value theory rewards player i with the
4) Apply the greedy-knapsack resource allocation strategy
Shapley value payoff φi (vE ) according to its contribution to
to select an optimal set of bearers to allocate the resource
resource sharing, which is based on the following four basic
to each individual QCI class.
axioms [23].
5) Define how much data a bearer can transmit in this time
1) Efficiency axiom: All  the resource blocks are completely slot.
distributed, that is, i∈P φi (vE ) = vE (P ).
After computation of QoS and data parameters, the GBR
2) Symmetry axiom: If for two players i1 and i2 , vE (S ∪
bearers’ data rate guarantees are met in the second step. Then,
{i1 }) = vE (S ∪ {i2 }), then for every coalition S ⊂ P
the remaining GBR classes’ resource quotas are allocated to
containing players i1 and i2 , φi 1 (vE ) = φi 2 (vE ).
their remaining bearers and the resource quotas of non-GBR
3) Dummy axiom: If for player i, vE (S) = vE (S ∪ {i}),
classes are allocated to their contained bearers for transmission
then for every coalition S not containing i, φi (vE ) = 0.
of their packets through the following steps. In the third step,
4) Marginality axiom: If τ and ϑ are two characteristic
every bearer b is assigned an overall rank value ρb , which is com-
functions and τ (S) − τ (S − i) = ϑ(S) − ϑ(S − i), then
puted by using a fine-tuned version of the multicriteria ranking
φi (τ ) = φi (ϑ) for every S ⊂ P .
function proposed in [22], aggregating the parameter-specific
Young [21] proved that the Shapely value is unique through
functions as follows:
the following theorem.

Theorem 1 (see [21]): The Shapley value leads to a unique ρb = f (mi , ωi ),
allocation solution to transfer utility games, satisfying effi- (9)
ciency, symmetry, and marginality. ∀i ∈ {delay, loss, queue depth, priority, throughput}.
Therefore, the Shapley value is considered to be a solution
of our coalitional game through which the equilibrium solution Each parameter-specific rank function f outputs a weighted
(i.e., behavior of the users in the steady state) can be obtained. In rank value, bounded in [0, ωi ], and calculated as
this approach, φi indicates the portion of resources that player
f (mi , ωi ) = ωi . tanh (mi ) (10)
i ∈ S receives from the division of vE (S) and is computed as
 (s − 1)!(p − s)!  where mi is the normalized value of parameter i. For the QoS
φi (vE ) = vE (S) − vE (S − {i}) parameters, mi is calculated by
p!
S ⊂P ,i∈S
(8) measured value of parameter i
where s = |S| indicates the number of players in coalition S. mi = (11)
QoS constraint of parameter i
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERDOSIAN et al.: FAIR-QoS BROKER ALGORITHM FOR OVERLOAD-STATE DOWNLINK RESOURCE SCHEDULING IN LTE NETWORKS 7

ωi is the adjustable weight for that parameter, which enables the

system designers to emphasis each particular parameter.
The computed bearer’s rank value represents the value im-
posed on the system by allocating resources to that bearer;
therefore, the objective of maximizing total system performance
is fulfilled by selecting the most valuable bearer items. Owing
to the fact that bearers with different values demand different
amounts of radio resources, the optimal solution is achieved by
using the greedy-knapsack solution, which iteratively selects the
bearers with a better Bearer Rank Value/Bearer Size ratio.
In the last step of QLA, resource quotas defined for each
QCI class are distributed to that class’s bearers in proportion
to their corresponding rank value. Therefore, the number of
resource blocks appropriated for the bearer bk from QCI class
k is computed as


ρb k k
Υb k = n k φ (13)
j =1 ρj

where nk is the total number of bearers from QCI class k waiting

for scheduling and φk denotes the whole number of unallocated
resource blocks dedicated to this class. In this way, the selected
optimum bearers transmit their data traffic over their respective
defined resource share in the network.
Regarding the complexity of the QLA algorithm, the while
loop requires O(|N |) iterations. In each iteration of the while
loop, the bearer with the highest rank value is searched. Sorting
the bearers of N in the decreasing order makes the pointer to the
highest ranked bearer immediately available in each iteration.
Since the sorting is executed before the while loop, and usually
requires O(|N | log |N |) iterations, then the complexity of the
QLA algorithm is O(|N | + |N | log |N |). Therefore, due to the
linear/logarithmic complexity, the proposed approach can be
considered to be scalable with the number of users or bearers.

V. SIMULATION ENVIRONMENT AND PARAMETERS

In this study, the simulation platform used to evaluate the per-
formance of the proposed multilevel scheduling algorithm FQB
is the same simulation platform applied in [18]. This simulation
platform was implemented based on the LTE network char-
acteristics defined in 3GPP LTE verification framework [24],
comprehending scheduling aspects of eNodeB MAC layer. This
simulation implementation allows us to evaluate the downlink
scheduling performance with different sources of traffic under
various simulation scenarios. The important simulation param-
eters are listed in Table II. The simulation setup consists of an
LTE network, based on EUTRAN architecture, consisting of
and for the throughput parameter, the normalized value of
eNodeB base station and mobile devices located in coverage
throughput is computed by
area of eNodeB. The implemented eNodeB provides the air in-
Estimated throughput terface as the communication links to the users. The voice and
mthroughput = . (12) data traffic were modeled by an exponential distribution func-
Past average throughput
tion and aggregate self-similar pattern [25], respectively, so that
The past average throughput is the data rate history of each they were realistic models, particularly in the overload state.
bearer and is updated every TTI when the bearer is allocated a re- The assumed 20-MHz bandwidth consists of 100 resource
source. The estimated throughput is the supportable data rate of blocks per spectrum allocation in the time–frequency domain.
each bearer, which is estimated by utilizing the channel quality A mixed traffic scenario composed of a stochastic distribution
indicator value of the users to which the bearer belongs. Weight of normal and overload time-intervals was imposed in which
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE SYSTEMS JOURNAL

TABLE II each QCI class is defined as the fraction of demand, as

SIMULATION PARAMETERS
ni (Δt)/li (Δt), if ni (Δt) < li (Δt)
xi (Δt) = (15)
Parameter Value 1, otherwise
System bandwidth 20 MHz
Frame structure TDD
where ni (Δt) and li (Δt), respectively, represent the time-
TTI 1 ms average throughput and total offered load perceived over the
Simulation length 32 min ith time interval by the whole given class.
Scheduling time window 10 ms
Number of data bearers 400
Hurst parameter for data traffic 0.9 B. Per-Class Fairness Provisioning
Number of voice bearers 1200
Voice activity factor for voice generator 0.5 To illustrate the relative equal share of the resources among
Mean talk spurt duration 5s the bearers inside each class, we evaluate the fairness for each
Voice codec encoding frame size 20 ms
Subcarrier spacing 15 kHz
QCI class separately by computing the cumulative distribution
Modulation and coding schemes QPSK, 16 quadratic-amplitude function of the per-class Jain’s fairness index during the simu-
modulation (QAM) and 64 QAM lation window time, as shown in Figs. 4 and 5. Based on these
Weight of throughput, loss, delay, queue depth 4, 4, 16, 4, 2
and priority parameters in order figures, we can deduce that the level of fairness obtained by
Run intervals (1, 5, 3, 2, 1, 2, 1, 10, 3, 5 and 3) the FQB algorithm is better than that obtained by the reference
scheduling algorithms for overload states.
there was a mixture of different traffic types. During the normal For QCI class 1 in Fig. 4(a), the fairness index value obtained
period, there was a set of active users in a cell with different kinds by FQB is between 0.88 and 1.00, which is the biggest range
and numbers of bearers (single VoIP, single data, multiple data, obtained by the algorithms, indicating that the FQB algorithm
and VoIP-data). For implementing the overload time interval, provides an optimal quality of experience (QoE) for VoIP bear-
200 users with single data session were added to the existing ers, even when there is an excessive load on the system.
traffic and they were eliminated at the beginning of the normal In case of QCI class 2 in Fig. 4(b), 50% of the bearers re-
state. ceived a fairness of 0.75 or less when FQB was used, while they
received a fairness of 0.65 or less when the proportional algo-
rithm was used. QCI class 3 has a latency/delay requirement of
VI. PERFORMANCE EVALUATION
50 ms, which makes it the most critical QCI class in terms of
In this section, the performance of the proposed scheme is latency. As Fig. 4(c) shows, the FQB algorithm obtained the best
evaluated with respect to the fairness, QoS, and throughput pa- fairness value for QCI class 3 compared to the other algorithms,
rameters. Inasmuch as no scheduling algorithm will be optimal which all produced a smaller amount of fairness.
for all service classes under overload conditions in the network, In contrast, for the non-GBR QCI classes, as can be seen
here we explore the tradeoffs of algorithms for various QCI in Fig. 5(a), (c), and (d), the difference between the FQB and
classes using several graphical and tabular results. alternative scheduling algorithms is less. For instance, the FQB
algorithm shows a better level of fairness for QCI class 6, where
A. Per-Class Fairness Evaluation Model 70% of the bearers from QCI class 6 experience a fairness of up
to 0.77, whereas for the knapsack and proportional algorithms
There are several opinions as to what is the best fairness
they experience fairness of 0.70.
achievement. Most of the scheduling algorithms proposed in
Overall, we can see that FQB outperforms the proportional
the literature employ two well-known quantitative fairness mea-
algorithm for almost all QCI classes, with the highest amount of
sures IEEE 802.16m [26] and Jain’s fairness index [27] to assess
improvements in contrast to the knapsack and greedy-knapsack
fairness in terms of the gained throughput.
algorithms. For the knapsack and greedy-knapsack algorithms,
Even though an uneven throughput division among the bear-
the fairness results for most of the QCI classes are very close.
ers of different QCI classes associated with various constraints
For the most part, the greedy-knapsack algorithm benefits from
and requirements is inevitable, a strict computation of fairness
the throughput-aware ranking function and produces better re-
just based on the actual gained throughput is not accurate for
sults than the knapsack algorithm, especially for QCI classes 1
multiservice scenarios. In this regard, a fairness verification
and 9.
model suitable for multiservice systems is per-class fairness,
The FQB improvements essentially benefit from the coopera-
which assesses the fairness criteria among the bearers within
tive coalitional game technique applied to resource distribution
the same QCI class by using Jain’s fairness index, defined as
at both the nesting and nested game levels, where the QCI

[ ti=1 xi (Δt)]2 classes, with different characteristics and different amounts of
fx (Δt) = t (14) resource demand, negotiate, and cooperate with each other to
t i=1 xi (Δt)2
make a fair decision. In contrast, in the reference algorithms,
where xi (Δt) is the metric of allocation during the ith time the resource competition among different service classes is ac-
interval (Δt = 1000 TTI). Because each time interval contains complished in a selfish manner, where each class’s performance
a different number of bearers demanding different amounts of desire is in conflict with that of the others. Therefore, by com-
resources, the metric of allocation in Jain’s fairness index for bining the results of Figs. 4 and 5, it can be concluded that the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERDOSIAN et al.: FAIR-QoS BROKER ALGORITHM FOR OVERLOAD-STATE DOWNLINK RESOURCE SCHEDULING IN LTE NETWORKS 9

Fig. 4. Per-class fairness for bearers from GBR classes: (a) QCI class 1, (b) QCI class 2, (c) QCI class 3, and (d) QCI class 4.

Fig. 5. Per-class fairness for bearers from non-GBR classes: (a) QCI class 6, (b) QCI class 7, (c) QCI class 8, and (d) QCI class 9.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE SYSTEMS JOURNAL

TABLE III TABLE VII

QoS AND THROUGHPUT EVALUATION, QCI1 QoS AND THROUGHPUT EVALUATION, QCI6

Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s) Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s)

FQB 21.34 4.5 0.00 FQB 4.29 3217.4 43.35

Greedy knapsack 21.23 4.5 0.00 Greedy knapsack 5.97 1386.4 20.45
Improved percentage (%) 0.5 0.0 0.0 Improved percentage (%) −28.1 −132.1 −112.0
Knapsack 21.14 4.5 0.00 Knapsack 5.83 1,423.1 18.69
Improved percentage (%) 0.9 0.0 0.0 Improved percentage (%) −26.4 −126.1 −131.9
Proportional 21.05 4.5 0.00 Proportional 5.97 1537.1 25.22
Improved percentage (%) 1.4 0.0 0.0 Improved percentage (%) −28.1 −109.3 −71.9

TABLE IV TABLE VIII

QoS AND THROUGHPUT EVALUATION, QCI2 QoS AND THROUGHPUT EVALUATION, QCI7

Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s) Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s)

FQB 12.15 26.1 0.00 FQB 3.78 1,203.1 11.62

Greedy knapsack 12.09 65.0 0.09 Greedy knapsack 5.02 612.6 4.02
Improved percentage (%) 0.5 59.8 100.0 Improved percentage (%) −24.7 −96.4 −189.0
Knapsack 11.84 68.5 0.17 Knapsack 4.74 550.3 3.41
Improved percentage (%) 2.6 61.9 100.0 Improved percentage (%) −20.2 −118.6 −240.8
Proportional 12.27 74.6 0.15 Proportional 4.60 923.0 6.64
Improved percentage (%) −1.0 65.0 100.0 Improved percentage (%) −17.8 −30.3 −75.0

TABLE V TABLE IX
QoS AND THROUGHPUT EVALUATION, QCI3 QoS AND THROUGHPUT EVALUATION, QCI8

Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s) Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s)

FQB 18.74 31.4 0.01 FQB 4.90 4,012.3 56.92

Greedy knapsack 18.62 52.1 0.06 Greedy knapsack 6.64 1,861.9 34.66
Improved percentage (%) 0.6 39.7 83.3 Improved percentage (%) −26.2 −115.5 −64.2
Knapsack 18.96 63.4 0.10 Knapsack 6.62 1,887.1 33.49
Improved percentage (%) −1.2 50.5 90.0 Improved percentage (%) −26.0 −112.6 −70.0
Proportional 18.73 70.7 0.18 Proportional 6.58 2898.5 54.16
Improved percentage (%) 1.0 55.6 94.4 Improved percentage (%) −25.5 −38.4 −5.1

TABLE VI
QoS AND THROUGHPUT EVALUATION, QCI4 TABLE X
QoS AND THROUGHPUT EVALUATION, QCI9

Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s)

Scheduler Avg. Throughput Avg. Latency (ms) Avg. Loss (Mb/s)
FQB 10.57 29.5 0.01
FQB 4.81 3903.3 56.11
Greedy knapsack 10.67 68.3 0.14
Greedy knapsack 5.55 2582.0 42.89
Improved percentage (%) −0.9 56.8 92.8
Improved percentage (%) −13.3 −51.2 −30.8
Knapsack 10.16 70.9 0.19
Knapsack 5.71 2537.0 39.65
Improved percentage (%) 4.0 58.4 94.7
Improved percentage (%) −15.8 −53.8 −41.5
Proportional 9.84 78.3 0.19
Proportional 4.67 3921.3 67.70
Improved percentage (%) 7.4 62.3 94.7
Improved percentage (%) 3.0 0.4 17.1

division rule applied in the FQB algorithm obtains an outcome

obtained by the FQB algorithm with respect to the reference
with fair service and ubiquitous coverage. Consequently, the
algorithms for the bearers of that class. The average through-
throughput of the system is not biased in favor of specific kinds
put, packet loss, and delay for GBR classes 1–4, presented in
of services.
Tables III–VI, indicate that the proposed algorithm performs
well enough to ensure that GBR QCI classes 1–4 meet their
C. Per-Class QoS and Throughput Enhancement
QoS constraints in terms of loss and delay, leading to a strong
The QoS performance results are presented in terms of av- QoE for GBR traffic.
erage packet loss and delay in addition to throughput for all As can be seen in these tables, the FQB algorithm obtains
QCI classes through Tables III–X. Each table is assigned to an improvements for GBR application classes, especially for QCI
individual QCI class and shows the percentage of improvement class 2. Moreover, the VoIP bearers, which correlate to QCI
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

FERDOSIAN et al.: FAIR-QoS BROKER ALGORITHM FOR OVERLOAD-STATE DOWNLINK RESOURCE SCHEDULING IN LTE NETWORKS 11

class 1, are scheduled with no loss and almost no delay. The performance targets, in terms of fairness assurance, QoS pro-
conversational video traffic from QCI class 2 and the rest of the visioning, and throughput maximization, without having to
GBR bearers from QCI classes 3 and 4 also experience near modify LTE specifications. The FQB algorithm is resilient in
zero loss. In addition, the non-GBR traffic for QCI classes 6–9 overload states when the users’ resource demands exceed the
experience reasonable levels of loss and delay, although they available system resources. A resource distribution approach
are impacted by the overload. was developed by applying the game theory model in a nested
Real-time applications, especially video and VoIP, which manner to provide per-class fairness among all service classes.
form the main volume of the current network traffic and are This proposed resource-to-class distribution scheme was fine
growing explosively, are sensitive to delay and need to be served tailored with the greedy knapsack to produce an efficient re-
with a guaranteed bitrate. Therefore, the proposed resource-to- source scheduling algorithm. Simulation results demonstrate
class distribution scheme in the first level of the FQB algorithm that queue-, QoS- and throughput-aware scheduling policies
grants more resources to the GBR bearers at the expense of low should be used in an LTE downlink scheduling mechanism
priority non-GBR bearers. to guarantee throughput and QoS requirements of the diverse
With respect to throughput performance, the FQB algorithm classes of traffic and provide fairness concurrently.
achieves a less notable throughput improvement (for example,
the same throughput in QCI classes 2 and 6) with respect to the ACKNOWLEDGMENT
greedy-knapsack scheduling algorithm because both of them The authors would like to thank Dr. M. Brehm and
use the same throughput-aware policy, which is applied in an Prof. Dr. R. Prakash for their valuable contributions from the
aggregate ranking function. Moreover, the better improvement University of Texas at Dallas.
in throughput acquired by FQB with respect to the proportional
and knapsack algorithms indicates that these algorithms are not REFERENCES
throughput optimal. The GBR service bearers, especially those [1] E. Dahlman, S. Parkvall, and J. Skold, 4G: LTE/LTE-Advanced for Mobile
from QCI classes 1 and 3, maintained a high level of QoE in Broadband, 2nd ed. New York, NY, USA: Academic, 2013.
[2] M. Aslam and M. N. Ayyaz, “Real-time delivery of 4G services with cross-
terms of throughput in the FQB scheduling algorithm. However, layered and power-optimized cognitive radio architecture,” IEEE Syst. J.,
some tradeoffs were also perceived. For instance, QCI classes vol. 10, no. 1, pp. 325–334, Mar. 2016.
6–8 were allocated less throughput, but the service traffic did [3] T. O. Olwal, K. Djouani, and A. M. Kurien, “A survey of resource manage-
ment towards 5G radio access networks,” IEEE Commun. Surveys Tuts.,
not starve, even under overload states in the network. Everyday vol. 18, no. 3, pp. 1656–1686, Thirdquarter 2016.
TCP traffic, which is assigned to QCI class 9, experienced a [4] Y. Liu, S. Xie, R. Yu, Y. Zhang, and C. Yuen, “An efficient MAC pro-
significant improvement in throughput compared to the results tocol with selective grouping and cooperative sensing in cognitive radio
networks,” IEEE Trans. Veh. Technol., vol 62, no. 8, pp. 3928–3941,
of the knapsack scheme. Oct. 2013.
The FQB’s strong bias toward providing fairness for all [5] N. U. Hassan, C. Yuen, S. Saeed, and Z. Zhang, “Power control for sum-
classes of QCI excessively compromised the experienced rate maximization on interference channels under sum power constraint,”
IEEE Trans. Veh. Technol., vol. 64, no. 2, pp. 593–609. Feb. 2015.
throughput and QoS over the region of interest (GBR classes) [6] Y. Liu, C. Yuen, X. Cao, N. U. Hassan, and J. Chen, “Design of a scalable
and lower priority services (non-GBR classes). Consequently, hybrid MAC protocol for heterogeneous M2M networks,” IEEE Internet
the results of the FQB algorithm indicate that this scheduling Things J., vol. 1, no. 1, pp. 99–111, Feb. 2014.
[7] F. Capozzi, G. Piro, L. A. Grieco, G. Boggia, and P. Camarda, “Downlink
strategy is the most effective for fine-tuning performance targets packet scheduling in LTE cellular networks: Key design issues and a
across the various service classes. survey,” IEEE Commun. Surveys Tuts., vol. 15, no. 2, pp. 678–700, Second
For the future research, the system model might be improved Quarter 2013.
[8] M. Carlesso, A. Antonopoulos, F. Granelli, and C. Verikoukis, “Uplink
to provide a balanced resource allocation among users for their scheduling for smart metering and real-time traffic coexistence in LTE
different traffic loads, while managing to be overall satisfying. It networks,” in Proc. IEEE Int. Conf. Commun., 2015, pp. 820–825.
would also be interesting to investigate the proposed scheduling [9] K. Ivesic, L. Skorin-Kapov, and M. Matijasevic, “Cross-layer QoE-driven
admission control and resource allocation for adaptive multimedia services
approach through multioperator shared systems to be adopted in LTE,” J. Netw. Comput. Appl., vol. 46, pp. 336–351, 2014.
in future heterogeneous networks. [10] G. Piro, L. A. Grieco, G. Boggia, R. Fortuna, and P. Camarda, “Two-level
downlink scheduling for real-time multimedia services in LTE networks,”
IEEE Trans. Multimedia, vol. 13, no. 5, pp. 1052–1065, Oct. 2011.
[11] N. Zorba and C. Verikoukis, “Energy optimization for bidirectional mul-
VII. CONCLUSION timedia communication in unsynchronized TDD systems,” IEEE Syst. J.,
vol 10, no. 2, pp. 797–804, Jun. 2016.
The development of the LTE technology to achieve system [12] M. E. Aydin, R. Kwan, and J. Wu, “Multiuser scheduling on the LTE down-
performance targets has arisen a number of challenges to the link with meta-heuristic approaches,” Phys. Commun., vol. 9, pp. 257–265,
capability of radio base stations for managing the bandwidth 2013.
[13] M. Kalil, A. Shami, and A. Al-Dweik, “QoS-aware power-efficient sched-
resources. In this regard, an efficient resource scheduling solu- uler for LTE uplink,” IEEE Trans. Mobile Comput., vol. 14, no. 8,
tion is required to keep in check the LTE design performance pp. 1672–1685, Aug. 2015.
targets. Furthermore, it is more desirable to design a resource [14] M. Al-Rawi, R. Jantti, J. Torsner, and M. Sagfors, “On the performance
of heuristic opportunistic scheduling in the uplink of 3G LTE networks,”
scheduling algorithm that is applicable for all service classes in Proc. 19th Int. Symp. Pers., Indoor Mobile. Radio Commun., 2008,
and considers heavily loaded states of the network, in contrast pp. 1–6.
to the existing service-specific algorithms. [15] A. Bousia, E. Kartsakli, A. Antonopoulos, L. Alonso, and C. Verik-
oukis, “Game-theoretic infrastructure sharing in multioperator cellular
In this paper, a multiservice resource scheduling algorithm, networks,” IEEE Trans. Veh. Technol., vol. 65, no. 5, pp. 3326–3341,
called FQB, was proposed to provide a tradeoff among LTE May 2016.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE SYSTEMS JOURNAL

[16] F. Shams, G. Bacci, and M. Luise, “An OFDMA resource allocation [23] Z. Han, D. Niyato, W. Saad, T. Basar, and A. Hjorungnes, Game Theory
algorithm based on coalitional games,” EURASIP J. Wireless Commun. in Wireless and Communication Networks: Theory, Models, and Applica-
Netw., vol. 1, pp. 1–13, 2011. tions. Cambridge, U.K.: Cambridge Univ. Press, 2012.
[17] E. Yaacoub and Z. Dawy, “A game theoretical formulation for proportional [24] 3GPP TSG-RAN1 R1–070674, “LTE physical layer framework for per-
fairness in LTE uplink scheduling,” in Proc. Wireless Commun. Netw. formance verification,” 3GPP TSG RAN1 #48, 2007.
Conf., 2009, pp. 1–5. [25] G. Kramer,“On generating self-similar traffic using pseudo-pareto distri-
[18] M. Brehm and R. Prakash, “Overload-state downlink resource allocation bution,” Technical brief, Dept. Comput. Sci., Univ. California at Davis,
in LTE MAC layer,” Wireless Netw., vol. 19, no. 5, pp. 913–931, 2013. Davis, CA, USA, 2000.
[19] N. Ferdosian, M. Othman, B. M. Ali, and K. Y. Lun, “Greedy-knapsack [26] IEEE C802.16m-07/306, “Protocol structure to support cooperative trans-
algorithm for optimal downlink resource allocation in LTE networks,” mission,” IEEE 802.16 Broadband Wireless Access Working Group.
Wireless Netw., vol. 22, no. 5, pp. 1427–1440, 2016. [27] R. Jain, The Art of Computer Systems Performance Analysis: Techniques
[20] I. J. Curiel, M. Maschler, and S. H. Tijs, “Bankruptcy games,”Z. Fur Oper. for Experimental Design, Measurement, Simulation and Modeling, 1st ed.
Res., vol. 31, no. 5, pp. A143–A159, 1987. New York, NY, USA: Wiley, 1991.
[21] H. P. Young, “Monotonic solutions of cooperative games,” Int. J. Game
Theory, vol. 14, no. 2, pp. 65–72, 1986.
[22] N. Ferdosian, M. Othman, B. M. Ali, and K. Y. Lun, “Throughput-aware
resource allocation for QoS classes in LTE networks,” Procedia Comput.
Sci., vol. 59, pp. 115–122, 2015. Author photographs and biographies not available at the time of publication.