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 COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION 1

A Survey on Opportunistic Scheduling in

Wireless Communications
Arash Asadi, Student Member, IEEE, and Vincenzo Mancuso, Member, IEEE

Abstract—Wireless technology advancements made oppor- implementations of schedulers do not utilize the advanced
tunistic scheduling a popular topic in recent times. However, features of physical layer and users are often scheduled
opportunistic schedulers for wireless systems have been studied regardless of their channel conditions. On one hand, fourth
since nearly twenty years, but not implemented in real systems
due to their high complexity and hardly achievable requirements. generation cellular technologies mostly operate on multiple
In contrast, today’s popularity of opportunistic schedulers ex- subcarriers (e.g., OFDMA). On the other hand, the majority
tends to implementation proposals for next generation cellular of newly proposed schedulers require user channel information
technologies. Motivated by such a novel interest towards op- to perform scheduling and it is not practical to feedback every
portunistic scheduling, we provide a taxonomy for opportunistic user’s channel condition over all subcarriers.
schedulers, which is based on scheduling design’s objectives;
accordingly, we provide an extensive review of opportunistic As a consequence, commonly adopted schedulers operate
scheduling proposals which have appeared in the literature regardless of user’s channel condition, e.g.: a) Round Robin,
during nearly two decades. that serves users in a circular manner without any other
The huge number of papers available in the literature propose consideration, b) earliest deadline first (EDF), which schedules
different techniques to perform opportunistic scheduling, rang-
the packet that will be expired the soonest, c) weighted fair
ing from simple heuristic algorithms to complex mathematical
models. Some proposals are only designed to increase the total queueing (WFQ), that allocates the resources with respect to
network capacity, while others enhance QoS objectives such the weights associated with every user [6]. In contrast, sched-
as throughput and fairness. Interestingly, our survey helps to ulers have been proposed, but not fully implemented yet, that
unveil two major issues: (i) the research in opportunistic is take advantage of physical layer information, such as the user
mature enough to jump from pure theory to implementation, channel state. This type of schedulers are called opportunistic
and (ii) there are still under-explored and interesting research
areas in opportunistic scheduling, e.g., opportunistic offloading of schedulers. As an example, some vendors deploy simplified
cellular traffic to 802.11-like networks, or cooperative/distributed opportunistic schedulers which follow the proportional fair
opportunistic scheduling. strategy [7]. Proportional fair schedulers consider the current
Index Terms—Resource allocation, Opportunistic Scheduling. channel state of the users and the history of received through-
put of each user. Unfortunately, the implementation accuracy
of proportional fair schedulers is commonly limited by scarce
I. I NTRODUCTION memory and processing resources deployed at base station.

B ROADBAND data provisioning became a design prin-

ciple for 3G and 4G networks as the dominant load on
cellular network has switched from voice to data traffic. The
The recent literature does not offer any survey on oppor-
tunistic scheduling, discussing the state-of-the-art proposals
and approaches in this field. Therefore, given the importance
emerging wireless technologies (e.g. HSPA [1], LTE [2] and of opportunistic scheduling in today’s communications, our
WiMAX [3]) have sophisticated multiple access methods such aim is to provide an extensive survey on the opportunistic
as orthogonal frequency division multiple access (OFDMA) scheduling algorithms emerged within the last two decades.
and code division multiple access (CDMA). In spite of capac- We start the survey with a brief introduction to opportunistic
ity increments created by new technologies, mobile operators scheduling. Next, the existing work is categorized based on
are still struggling to satisfy users demands [4], [5]. A viable our proposed taxonomy, namely based on capacity, fairness,
alternative to bandwidth increment is to allocate the existing QoS issues, and distributed / centralized assumptions, and the
bandwidth optimally or at least more efficiently. Schedul- significant works under each category are reviewed. In order
ing is the functionality responsible for resource allocation to provide a better understanding of the proposals, we include
among users, and its job is to decide which user should a comparison and evaluation of such proposals in the end
transmit/receive and when, therefore it impacts on bandwidth of each section. A list of common techniques, approaches
utility efficiency. and evaluation methods is provided based on the reviewed
In today’s cellular networks, the design and choice of literature. In addition, we discuss open research problems
schedulers is left to the operator. As a consequence, the with some suggested solutions and possible future direction
significant impact of scheduling algorithms on network per- of research on opportunistic scheduling.
formance made them a popular research topic. The current The rest of the paper is organized as follows: in Section II,
we give a short introduction on opportunistic scheduling
Manuscript received April 25, 2012; revised September 25, 2012. algorithms, considering their approach through opportunism.
The authors are with Institute IMDEA Networks, Madrid, Spain and Carlos
III University, Madrid (email: {arash.asadi, vincenzo.mancuso}@imdea.org). We also provide a taxonomy for opportunistic schedulers.
Digital Object Identifier 10.1109/SURV.2013.011413.00082 In Section III, we review schedulers that increase the total
1553-877X/13/$31.00 
c 2013 IEEE
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 COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

TABLE I
T HE SCHEDULING POLICY FOR M AX R ATE , M AX W EIGHT AND E XP RULE SCHEDULERS .
Scheduler Scheduling policy Notations
MaxRate i∗ (t) = arg max Ri (t) N : Total number of users.
i∈N
MaxWeight i∗ (t) = arg max Ri (t)Qi (t) i∗ (t): The user which is scheduled at time t.
i∈N
ai Qi (t)
Exp rule i∗ (t) = arg max γi Ri (t)e β+[Q̄(t)]η Ri (t): Achievable rate by user i at time t.
i∈N
Qi (t): Queue size of user i at time t.
γi , ai : Arbitrary positive fixed constants.
η,β ∈ (0, 1). 
.
Q̄(t) = (1/N ) i ai Qi (t)

network capacity. In Section IV, we survey QoS-oriented Opportunistic scheduling has been proposed not only to
proposals. Section V contains proposals with fairness consid- improve capacity or QoS. For instance, Wong et al. proposed
erations. State of the art distributed opportunistic schedulers an opportunistic scheduling strategy to leverage multiuser
are discussed in Section VI. Some of the open issues on diversity in an OFDM systems and do attempt to minimize
opportunistic scheduling are listed in Section VII. Finally, in the overall transmission power [16].
Section VIII, we conclude the paper and provide a list of Although opportunistic schedulers are not widely imple-
common techniques and evaluation approaches in opportunis- mented in commercial deployments yet, they have been proven
tic proposals. We also point out future research trends. to be more suitable than non-opportunistic schedulers for wire-
less networks. Initial proposals employed heuristics to make
II. O PPORTUNISTIC S CHEDULING scheduling decision which was improvement in comparison to
Here, we provide a brief introduction on opportunistic non-opportunistic schedulers but not optimal [17], [18]. For
scheduling. We point out incentives for employing opportunis- instance, Badia et al. propose a joint scheduling and resource
tic schedulers in cellular networks and provide a taxonomy for allocation framework tailored for WiMAX and operates based
the classification of opportunistic schedulers. on a heuristic credit-based scheduler [17]. Later, researchers
obtained mathematical models for wireless channels which al-
lowed them to study different aspect of the system (e.g., queue
A. Incentives behind opportunistic scheduling
stability, delay, throughput) and propose optimal schedulers.
Opportunistic schedulers take into account information The majority of recent literature is dedicated to modeling
such as the channel quality in terms of QoS metrics (i.e., and optimizing proposals for cellular networks using different
throughput, delay, jitter) that allows the scheduler to find techniques which will be discussed later. To explore these
the proper transmission resources for each user. The notion proposals, we now introduce our taxonomy for opportunistic
of opportunistic scheduling was first introduced by Knopp scheduling.
and Humblet in [8]. They showed that using the mutiluser
diversity in scheduling process can significantly improve the
capacity. In a pure opportunistic approach, the scheduler
always chooses the user in the best channel condition to B. Taxonomy
use the resources. This approach is referred to as MaxRate The available literature on opportunistic scheduling tack-
scheduling in the literature [9]. The gain in opportunistic les the issue of scheduling from different aspects. Most
scheduling depends on the multiuser diversity due to random of these proposals are subclasses of four major categories:
wireless channel impairments such as fading and multipath. capacity, QoS, fairness, and distributed scheduling. Propos-
After [8], researchers aimed at taking advantage of diversity als that purely improve network capacity regardless of QoS
caused by the channel impairments instead of eliminating it. or fairness implications are listed under the first category.
Some authors even propose techniques such as opportunistic We further categorize these proposals into schedulers with
beamforming to increase the multiuser diversity [10], [11], full/non-full channel state information and schedulers for cog-
[12]. With this technique, the same signal is transmitted nitive radio networks. The second category covers proposals
over multiple antennas with different transmission powers. that aim to improve specific quality indices such as delay,
This increases the channel diversity of users, which leads jitter, average throughput, etc. The works under this category
to improved opportunistic gain. MaxWeight [13] is another are also divided into subcategories of single QoS objective
opportunistic scheduler that selects the user with the highest and multiple QoS objective. Works under the third category
product of queue length and transmission rate. MaxWeight was tackle the fairness issue in opportunistic scheduling. In fact,
considered throughput-optimal before the authors in [14] prove opportunistic scheduling of users can lead to highly unfair
otherwise under flow level dynamics. However, MaxWeight treatment toward different users. Therefore, fairness is always
is throughput-optimal with fully backlogged queues. Exp a concern in opportunistic scheduling because users with low
rule schedulers [15] are throughput-optimal schedulers that channel quality can be sacrificed due to the greedy nature
prioritize users based on an exponential formula using queue of pure opportunistic approaches. Eventually, most of the
size and transmission rate of every user. Table I shows the proposals focus on mechanisms that can be implemented at
scheduling policy of the aforementioned throughput-optimal the base station of a cellular network. However, distributed
schedulers. mechanisms have been proposed as well. We classify these
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 3

Fig. 1. Opportunistic Scheduler Classification.

distributed opportunistic algorithms in a separate class because maximizes the product of queue size and feasible rate for
they usually aim at a different network configuration. each user over all subcarriers. The second and third objective
In this paper, we review the existing literature based on the functions are NP-hard problems that account for the ignorance
proposed taxonomy, which is schematically depicted in Fig. 1. of MaxWeight algorithm towards users with small queue
and bad channel quality, as discussed in [14]. To serve this
III. C APACITY purpose, the second objective function prioritizes the users
In many proposals, opportunistic scheduling is employed as with small queues and the third objective function maximizes
a solution to enhance the total capacity of wireless networks. the negative drift of a Lyapunov function [27] (i.e., maximizes
As shown in Fig. 1, we further classify these proposals based the queue length variation in every slot). The authors propose
on the assumption of full channel state information (CSI) five algorithms based on the objective functions defined for
availability (i.e., base station has instantaneous knowledge of MaxWeight. The algorithms which are derived from second
all users’ CSI). Hence, the scheduler always knows which user and third objectives inherit the NP-hardness. The authors solve
has the best channel state at every time instant [19], [20], [21], the NP-hard algorithms via approximations and prove their
[22], [23], [24], [25], [26]. However, note that CSI of mobile stability. The simulation results showed that the algorithms
users is acquired via feedback in cellular networks, so that based on the second and third objective provide better per-
the perfect and instantaneous (full) knowledge of user’s CSI formance. They also show that the algorithms which optimize
is hardly practical in real deployments. To cope with this issue, the scheduling decision over all carriers instead of local carrier
many proposals leverage the knowledge of user’s channel sta- optimization outperform the rest.
tistical behavior, and CSI samples, rather than the knowledge
Liu et al. [20] state that throughput-optimal algorithms
of the exact and instantaneous CSI. Significantly different is
in single channel wireless networks are not necessar-
the case of opportunistic schedulers improving capacity by
ily throughput-optimal in multi-channel wireless networks.
using cognitive radio techniques. Indeed, with cognitive radio,
Hence, they propose a joint channel-assignment and workload
full or partial CSI can be available, but opportunism mainly
based scheduler (CA-WS), which is throughput-optimal in
focuses on which resources can be used and when, given that
multi-channel wireless networks. In [20], flows are classified
some other primary user has strictly priority access to the
into two groups, namely, transient and resident and every flow
wireless resources. Therefore, in the following we describe
is associated with a file transfer from source to destination. A
first those proposals relaying on full CSI availability, then
transient flow is a flow whose file is not fully buffered at the
proposals not using full CSI, and eventually proposals based
base station, while a resident flow has fully transmitted the file
on cognitive radio.
to the base station. Every flow starts from a source and crosses
the base station to reach the destination. Let’s assume that
A. Resource allocation with full channel state information all flow sources try to transmit a file to the base station. The
Here, we survey the work attempting to enhance wireless scheduler keeps track of the number of allocated time slots on
network throughput under the assumption of full CSI avail- each channel for every flow and the CA-WS algorithm chooses
ability at the base station. the best combination of channels and slots for every user to
In many wireless technologies, users can transmit over maximize the network capacity. Under CA-WS, the transient
more than one carrier. This ability extends the opportunistic flows are not scheduled until their status is changed to resident.
scheduling decision process to carrier allocation among users. This induces longer transmission delays and degrades the
Andrews and Zhang [19] tackle the problem of scheduling performance of CA-WS under low and medium traffic load. To
in a multi-carrier wireless system. Their paper is dedicated solve this issue, the authors propose a hybrid CA-WS scheme
to adapt the MaxWeight algorithm for multi-carrier scenarios which uses CA-WS to serve resident flows and MaxWeight to
for which they define three objective functions that emulate serve transient flows. In particular, in order to eliminate the
the MaxWeight behavior. The first objective function simply full CSI assumption, the authors propose to use the learning
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 COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

TABLE II
process that was earlier proposed in [28]. The performance T HE EFFECT OF PROPOSED ALGORITHMS ON CALL BLOCKING
of CA-WS is compared with the MaxWeight scheduler via PROBABILITY NORMALIZED OVER M AX W EIGHT ACHIEVED RATE UNDER
HIGH TRAFFIC LOAD .
simulation. The results confirm that CA-WS performance is
poor under low and medium traffic load. However, CA-WS Algorithm Call blocking
probability normalized
performs much better in presence of high traffic load. CA-WS to MaxWeight’s blocking
has higher blocking probability than MaxWeight and hybrid probability (%)
scheduler because it waits for the files to be fully transmitted. WSL ≈ 71
Delay-Based ≈ 14
The results also show that the hybrid scheduler can serve 20% CA-WSU ≈ -100
more traffic than MaxWeight and CA-WS while maintaining Hybrid CA-WSU ≈ 60
a zero blocking probability. The hybrid scheduler has better
delay performance than CA-WS and MaxWeight. MaxWeight
has lower delay than CA-WS when the traffic is low, but its of [28] also provide the necessary conditions for stability of
performance degrades drastically with traffic increment. a scheduler which is: a) the service allocated to each users
Al-Zubaidy et al. in [23] target scheduling in high-speed should not be less than what was requested if the service were
downlink packet access (HSDPA) networks. Modeling the supportable at all. b) the total airtime allocated to short-lived
behavior of HSDPA network is challenging because it uses and long-lived flows should less than or equal than the total
both time division multiplexing and code division multiplex- available time. The authors prove that WSL is throughput-
ing. A 2-state finite state Markov chain (FSMC) [29] is used optimal.
to model channel state transition probabilities. The HSDPA In the same paper, Liu et al. discuss the basic problem
is modeled by a 5-tuple (T, S, A, Pss (a), R(s, a)), where T of MaxWeight, e.g., a flow with small backlog may never be
is time, S and A represent the system and action space. served. A solution for this problem is to use the product of the
Pss (a) and R(s, a) are the system state and reward functions head of line delay (delay-based scheduling). However, in [28]
when the system is taking action a in state s. The reward it is shown that the delay-based scheduler is also not stable in
function can be used to tune between throughput and fairness. presence of short-lived flows. The authors conclude the paper
The authors introduce the optimal scheduling policy based on with a set of simulations to evaluate the performance of WSL,
their proposed model and solve the model with the dynamic MaxWeight, and delay-based scheduler. It is shown that WSL
programming technique [30]. Next, they study the behavior can sustain a zero blocking probability while admitting almost
of the optimal policy using a Markov decision process (MDP) 20% more traffic. WSL also shows better delay performance.
based on which they propose a heuristic scheduling policy The authors of [20] and [28] benchmark their proposals
with near optimal performance but lower complexity. The against the MaxWeight algorithm. In Table II, we can see how
performance of the optimal policy and the one of the heuristic these proposals affect the blocking probabilities with reference
are shown to be comparable using simulation. Round Robin to MaxWeight. Since CA-WS increases the waiting time of the
scheduler is also included in the simulation as a benchmark. transient flows, it increases the call blocking probability that
System throughput under light traffic load is the same for results in lower performance than MaxWeight (i.e., negative
Round Robin, optimal, and heuristic policies. The optimal gain). The rest of the proposed schedulers improve the call
policy performs 10% and 25% better than Round Robin under blocking probability over MaxWeight.
medium and high traffic load, respectively. Al-Zubaidy et Ouyang et al. propose a scheduling algorithm with near-
al. [23] investigate the impact of the number of codes per user, optimal performance in [31]. The paper tackles the issue
i.e., the code chunk granularity, on the scheduling policy. It discussed earlier: in practice the scheduler will know a user’s
is observed that higher code chunk granularity increases the CSI after it received a transmission feedback from the user,
average queue delay. Hence, policies with finer granularity are not before. The classic exploration vs. exploitation problem
more favorable. comes to picture under this situation: the scheduler has to
decide whether to wait for more information hoping to make a
better decision (future gain) or to transmit based on the current
B. Resource allocation without full channel state information information (immediate gain). Ouyang et al. model the fading
Here, we describe the proposals considering that base channel via a two-state Markov chain, high hi and low li states
stations do not have direct and instantaneous access to user’s representing the maximum allowable data rates in the lower
CSI, but they can periodically acquire CSI from mobile user’s and higher states, respectively, and propose a joint estimation
reports, see Fig. 1. These non-full-CSI proposals are inherently and scheduling process using a partially observable Markov
different for the ones surveyed in the previous subsection, and decision process (POMDP). We can observe the scheduling
therefore lead to substantially different mechanisms. process in the Fig. 2.
In [28], Liu et al. propose a throughput-optimal scheduler As shown in Fig. 2, a belief value πi is defined for every
that does not require any prior knowledge of channel state user i that states the probability that channel Ci is in state hi .
and user demands. This can be achieved using the so called The belief value πi is updated via the feedback from user i at
workload-base scheduling with learning (WSL). The authors the end of each transmission period. The scheduler picks users
define the flows that continuously inject traffic as long-lived in a manner to maximize the total throughput of the system.
and those with finite number of bits as short-lived. In order The authors of [31] state that this scheduling problem is a
to find the maximum possible data rate of short-lived flows, restless multi-armed bandit process (RMBP) [32], [33]. They
their data rate is monitored for a learning period. The authors use the Whittle’s index policy to solve the RMBP problem.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 5

As mentioned earlier, these feedbacks are crucial for the

base station to make opportunistic scheduling decisions that
result in significant capacity gain in the network. Ouyang et
al. [38] propose an algorithm to maintain the opportunistic
gain while reducing the feedback overhead. They propose an
opportunistic weight-based feedback allocation (WBF) scheme
that provides a fraction of full throughput region under the
MaxWeight scheduling, while limiting the number of feed-
backs on every channel. WBF prioritizes users according to
Fig. 2. Opportunistic scheduling with joint channel estimation and schedul- their queue size and transmission rates and it allocates more
ing [31]. feedbacks to probe the channels of users with larger queues
and higher transmission rates. It is shown via simulations that
WBF can achieve almost identical results as compared to a
If the Whittle’s index policy exists, the scheduler has near scenario where full CSI is send to the base station.
optimal performance. Using numerical evaluation, they show
that the indexing policy shows near optimal performance and
provide more than 90% gain over a randomized policy that C. Cognitive radio networks
schedules users randomly with uniform distribution. This subsection covers the opportunistic schedulers pro-
In a realistic scenario, in order to make opportunistic deci- posed for cognitive radio networks under the capacity cate-
sions, the base station needs to acquire the channel information gory. In a cognitive radio network, secondary users have trans-
from users. In 3G, each user has a dedicated channel to mission opportunity only if primary users are not transmitting.
feedback this information. On one hand, we can save the It is desirable to design a scheduling scheme that improves
costly wireless bandwidth by reducing the amount of feedback the service received by secondary users while minimizing the
channels. On the other hand, the opportunistic decisions are collision possibility between primary and secondary users. The
made based on the channel information of users that is cognitive network control scheduling algorithm (CNC) [21]
obtained via these feedbacks. Therefore, again we encounter maximizes the throughput of secondary users while bounding
the classic exploitation vs. exploration trade-off. the maximum number of collisions with primary users. The
Chaporkar et al. [34] address this trade-off and propose authors of [21] consider a network with primary and secondary
a scheduler with joint probing and transmission strategy. users. There is a dedicated channel for every primary user.
They show that throughput-optimal schedulers under full CSI CNC consists of a flow control policy and a resource allocation
assumption are also optimal when CSI is obtained via probing policy. The flow control policy takes into account the current
users. They used a Markov chain to model the queues, and the backlog of secondary users to decide whether a new packet
Lyapunov drift technique [35] to proof stability. The choice of must be admitted to the queue or not. A virtual collision
optimum number of feedback requests is a generalized optimal queue is introduced that monitors how much a primary user
stopping time problem [36] with an additional decision on experiences collisions more than a predefined threshold. Using
choosing the users that should send their feedback first. The Lyapunov drift and Lyapunov optimization [35], the authors
authors state that optimal solution to this problem can be prove that CNC maintains a worst case bound for both backlog
found under special cases, hence they propose an approximate queue and number of collisions. The authors also propose to
scheduling policy and probing and transmission strategy. The use CNC in a distributed manner by using greedy maximal
results of numerical simulations show the maximum system match scheduling [39], [40]. The distributed implementation
throughput using full CSI and probing under different probing supports any rate within 50% of the capacity region. The
time β. It can be observed that probing channel has significant simulations results [21] show that CNC can bound the total
impact on throughput. It is also observed that, if the number average congestion while delivering almost all the traffic.
of users is large enough, there exists a threshold above which In [22], Khalil et al. propose a cooperative scheduling
probing extra users is unnecessary. scheme for cognitive radio networks. In a classic cognitive
Jacko studies the value of information in opportunistic network, secondary users utilize the slots which are not used
scheduling in [37]. He models the wireless network using by primary users. In contrast, Khalil et al. consider a scenario
the Gilbert-Elliot model considering flow dynamics of the in which secondary users in good channel state help primary
network, and proposes an opportunistic scheduler based on an users in bad channel to increase the channel capacity. The
optimal job (flow) indexing policy. Jacko shows that the chan- secondary users can be rewarded immediately or in the long
nel information becomes valuable when queues are not fully term. Assume a primary user sends its data to a secondary
backlogged. He also concludes that user’s channel steady- user using a (1 − α) fraction of the time slot, 0 ≤ α ≤ 1. If
state information is enough to make near optimal scheduling the secondary user transmits the primary user’s data using a
decisions. fraction β of the time slot, α(1 − β) is the remaining time that
Emerging cellular technologies such as LTE and WiMAX is saved by relaying data to secondary user in good channel,
operate with multiple channels and multiple users. If every 0 ≤ β ≤ 1. Fig. 3 is a depiction of the cooperative cognitive
user sends feedbacks for every available channel at the base network operation proposed in [22]. With the immediate re-
station, a big chunk of bandwidth is dedicated to the feedbacks ward scheme, the secondary user transmits its own data during
which is an unwanted load for the operators infrastructures. the remainder of the time slot. With the long term scheme, 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.
6 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

and other considerations taken into account besides capacity

improvement.

IV. Q UALITY OF S ERVICE

With the recent advent of applications such as VoIP and

video conferencing, quality of service (QoS) gained popularity
in both research and industry. There are several QoS objectives
defined such as throughput, delay, jitter, packet loss, error rate,
latency and so on. Among the QoS metrics, the opportunistic
scheduling proposals pay more attention to delay and through-
put as depicted in Fig. 1 . Some of the proposed schedulers
may adopt a single QoS objective [41], [42], [43] while others
use multiple objectives [44], [45]. In the following, we review
the proposals with single QoS objective. Then we continue
with proposals taking multiple QoS objectives into account.

A. Single QoS objective

In this subsection, we focus on proposals with one single
Fig. 3. Cooperative cognitive radio transmission slot structure. Primary user QoS objective. Among available QoS objectives, delay and
sends its data to secondary user in a fraction of 1−α. Secondary user forwards
this traffic in αβ and in a fraction of α (1 − β), the secondary user sends its
average throughput are the most common used.
own traffic. In [41], Kim and de Veciana investigate the performance
of opportunistic scheduling with heterogenous traffic (i.e.,
QoS and best effort (BE) flows). They show that traffic
scheduler guarantees to allocate a portion of resources saved integration—i.e., the coexistence of QoS and BE traffic in
by means of cooperation to secondary users. The stability of the same network—deteriorates the performance of the system
the proposed algorithm is proved using the Lyapunov drift in terms of capacity, stability, and delay. This performance
and optimization techniques. The proposed scheduler is not anomaly was previously dealt with at packet-level in [46],
optimal but it can be pushed towards optimality at the cost of [47], [48].
longer average queues. Numerical results illustrate that Khalil Kim and de Veciana [41] studied the interaction of QoS and
et al.’s proposal improves the total network utility up to 5%, BE traffic at flow-level for the first time. They find necessary
while providing secondary users with non-zero utility. and sufficient stability conditions for the traffic integration
models that was previously provided using a 2-dimentional
D. Summary Markov chain in [49], [50], [51]. The proposed opportunistic
The authors of [19], [20], [23] focus on throughput optimal scheduler is designed in a way that QoS flows receive a
scheduler in presence of full CSI. The proposals in [19], fixed average throughput per slot. Other QoS objectives such
[20] tries to adapt MaxWeight to multi-carrier scenarios. as delay or jitter are not considered. The authors argue that
The algorithms proposed in [19] are based on generalization allocating an average throughput to QoS flows in every slot
of MaxWeight with optimization problems for multi-carrier reduces the chance of starvation in the long period. BE traffic
networks. Due to hardness of these optimization problems, the is modeled as finite file transfers using HTTP or FTP and its
authors investigate on bounded approximations. CA-WS and performance is evaluated through the average time needed to
hybrid CA-WS [20] are shown to be both throughput optimal, complete a file transfer.
which resolve the weakness of MaxWeight in presence of flow Additionally, the authors of [41] propose an opportunistic
dynamics. scheduling scheme that monitors the number of QoS and BE
In [28], [31], [34], the authors consider the effect of flows. In order to be able to guarantee the fixed average
feedback on system performance. For instance users’ CSI is throughput in every slot, the maximum number of QoS flows
predicted by a short observation period in WSL [28] or by is limited such that the total promised bandwidth remains less
using POMDP as in [31]. Chaporkar et al. propose a technique than the total available bandwidth. It should be noted that
which reduces the number of feedback while maintaining average channel quality of users affects the total capacity of
throughput optimality. the network and the maximum number of QoS flows. Kim
In [21], authors exploit the concept of virtual collision and de Veciana also propose a bandwidth borrowing/lending
queues to reduce the number collisions between secondary and scheme that allows QoS services to borrow bandwidth from
primary users. Khalil et al. proposes a scheduling scheme that BE services when required. Therefore, each QoS flow bor-
increases the total system throughput by exploiting secondary rows bandwidth from BE flows to maintain the promised
users with good channel conditions [22]. Table III shows each average throughput. Similarly, QoS flows can borrow their
proposal mainly focusing on capacity with details regarding extra bandwidth to BE flows. They show that integration of
the assumptions taken by the authors, analytical tools used for QoS and BE flows reduces the system capacity and leads to
the proposals, the scenario in which the proposal in applicable, the so called loss in opportunism phenomenon (33% capacity
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 7

TABLE III
S UMMARY OF PROPOSALS WITH MAIN FOCUS ON capacity
Proposal Assumptions Analytical tools Topology Other focus
Scheduler for multicarrier Full CSI Greedy algorithms Single cell (OFDMA) Fairness
wireless system [19] Not fully backlogged Dynamic programming Multi-carrier Stability
Traffic: Generic Linear programming Downlink
Lyapunov drift
Scheduler for multichannel Full CSI Markov chain Single cell (OFDM) Throughput-optimal
wireless systems with Not fully backlogged Lagrangian multipliers Multi-carrier Stability
flow-level dynamics [20] Traffic: Generic Lyapunov drift Downlink
Optimal scheduler Full CSI Finite state Markov chain Single cell (HSDPA) Throughput
for HSDPA networks [23] Not fully backlogged Dynamic programming Downlink Delay
Error free transmission Markov decision process Fairness
Traffic: Bernoulli
Throughput-optimal scheduler Non-full CSI Markov chain Single cell Throughput-optimal
that accounts for Not fully backlogged Lagrangian multipliers Downlink Stability
flow-level dynamics [28] Traffic: Flow level with Lyapunov drift
two class of flows
Joint channel estimation Non-full CSI Markov chain Single cell
and scheduler for Restless multi-armed bandit process Downlink
wireless networks [31] Partially observable Markov
decision process
Throughput-optimal Non-full CSI Markov decision process Single cell Throughput-optimal
scheduling with limited Not fully backlogged Lyapunov drift Downlink Stability
channel information [34] Traffic: Generic Optimal stopping theory
Optimal feedback allocation Non-full CSI Markov chain Single cell (FDD) Near throughput-optimal
in multichannel wireless Not fully backlogged Downlink Stability
networks [38] Traffic: Poisson Multi channel
Flow-level scheduler Non-full CSI Markov decision process Single cell Stability
for wireless networks [37] Not fully backlogged Gilbert-Elliot model Downlink
Traffic: Bernoulli Lagrangian multipliers
Dynamic programming
Opportunistic scheduler Non-full CSI Markov chain Multi-cell Throughput-optimal
for cognitive radio Not fully backlogged Lyapunov drift
networks [21] Traffic: Bernoulli Lyapunov optimization
Maximum weight match
Optimal scheduler Full CSI Lyapunov drift Single cell Stability
for cooperative cognitive Fully backlogged Lyapunov optimization
radio networks [22]

reduction in the example provided in [41]). This loss is due average queue length of users. Sadiq et al. also evaluate the
to QoS requirements of flows, which forces the opportunistic performance of Log rule, MaxWeight, and Exp rule sched-
scheduler to transmit packets of QoS flows, although that was ulers via simulation. They show that Log rule and Exp rule
not the opportunistic choice at that moment. QoS flows also exhibit comparable delay performance and perform better than
affect the delay experienced by BE flows. This effect magnifies MaxWeight under low traffic. In a high traffic scenario, users
with lower SNR, higher guaranteed bandwidth, and larger experience 20-80% less delay under Log rule, see Fig. 4. It
number of QoS flows. If QoS flows remain in the system is also shown that Log rule degrades more gracefully that
for a long time, BE flows are under-served until they have a the Exp rule and MaxWeight under high traffic load. In other
chance to recover, i.e., QoS sessions leave the system. This words, when the system is overloaded more users meet their
phenomenon is called local instability which is caused by QoS requirements under Log rule in comparison to Exp rule
the coexistence of QoS and BE flows. Kim and de Veciana and MaxWeight.
propose a call admission control (CAC) for BE flows which Some schedulers are designed to perform under worst case
solves the local instability issue. Using numerical evaluation, scenarios while others do not consider any specific network
they show that using CAC reduces the local instability and conditions. Sadiq et al. conclude that the schedulers which are
improves the delay for BE flows. optimized for overall system performance are more likely to
In [42], Sadiq et al. focus on the performance and ro- be robust to changes than those optimized for the worst case.
bustness of packet schedulers in wireless networks. They In [43], Neely proposes an opportunistic scheduling al-
prove that mean-delay-optimal policies show radial sum- gorithm with delay guarantees. He develops a novel virtual
rate-monotonicity (RSM) meaning that when users’ queue queue technique (i.e., the -persistent service queue) which
grow linearly, the scheduler serves users in a way that de- guarantees a worst case delay for each users. He further
emphasizes queue balancing in order to maximize system uses Lyapunov drift and optimization techniques to obtain
throughput. This behavior is in contrast with the behavior of a throughput-optimal scheduling algorithm that guarantees
known throughput optimal schedulers (e.g., MaxWeight [13], bounded worst-case delay. The proposed scheduler is com-
Exp rule [15]). In that paper, Log rule policy is introduced patible with both ergodic and possibly non-ergodic channel
as a new class of policies that are RSM and throughput and arrival settings. Moreover, it can be used for both single-
optimal. The authors use a Markov chain to obtain the queue hop and multi-hop scenarios. Finally, the author proves that
state transition probabilities which are used to minimize the the performance of the proposed algorithm is comparable to
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 COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

Fig. 5. QoS unity cube [45].

tions over time. The proposal performance is illustrated using

computer simulations in both stationary and non-stationary
channel conditions. In the simulations, it is assumed that the
queues are fully backlogged and users move with the speed
of 50 km/h invariably. Results show that the throughput of
the proposed algorithm is on average 30% higher than that of
Modified Largest Weighted Delay First (M-LWDF) [53]. M-
Fig. 4. Performance comparison among Log rule, EXP rule and LWDF is a heuristic that was originally designed for TDMA
MaxWeight(MW) scheduling algorithm. Mean packet delays are shown under systems. It selects users based on a simple metric, taking
three different arrival rate, namely low, medium and high [42].
into account both the current channel state and the head-of-
line packet delay. Unlike M-LWDF, packet drop rate of the
schedulers that have advance knowledge of channel variations AADTR-based algorithm remains the same with increasing
(i.e., full CSI). number of users.
In [37], that was briefly discussed in the previous section, In [45], throughput, delay and packet drop are the adopted
Jacko proposes an opportunistic scheduling scheme using a QoS objectives of the scheduler. The authors map the QoS
Whittle indexing policy that minimizes the transmission delay. objectives into a unity cube that represents the QoS state of
Jacko’s proposal is also throughput-optimal under not fully the flows, see Fig. 5. The delay and packet drop incurred by
backlogged conditions. each flow is normalized over the flow’s QoS requirements. The
achieved throughput is, on the other hand, normalized over the
desired throughput, and the inverse of the normalized through-
B. Multiple QoS objectives put is used as third QoS dimension. Hence, a normalized value
In this subsection, we list the proposals that intend to meet equal or less than one indicates that the flow is satisfied under
more than one QoS metric. that QoS objective. A flow that is mapped inside the unity
In opportunistic scheduling, it is common to observe that cube is satisfied for all its QoS objectives.
users with low channel quality frequently experience transmis- The QoS efficient work point is at (1,1,1) and the QoS-
sion rate fluctuations. These fluctuations result in larger queues deviation metric is defined based on the distance from the
and longer delays. Choi et al. proposed the AADTR metric current position of the QoS descriptor of the flow in the cube
to be able to measure and control these fluctuations and their to the efficient work point. A negative value represents an
resulting delays [44]. The algorithm proposed by Choi et al. unsatisfied flow and a positive value represents a satisfied
in [44] targets orthogonal frequency division multiple access flow. The authors propose three schedulers, namely, biggest
(OFDMA) wireless networks in which users can transmit QoS-deviation first (BQDF), adaptive QoS-deviation control
over different subcarriers at the same time. Their proposal (AQDC), and adaptive residual time control (ARTC). BQDF
maximizes system throughput while meeting the required chooses the user with the highest deviation that corresponds to
average transmission rate and the average absolute deviation the least satisfied flow. BQDF causes throughput degradation
of transmission rate (AADTR). The latter is a metric to because of ignoring channel condition of the flows. AQDC is
control the transmission rate fluctuations. QoS flows have both proposed to fix the biased behavior of BQDF. AQDC builds a
average transmission rate and AADTR objectives. Average set which includes the p% of flows with highest QoS-deviation
transmission rate is the only objective for BE flows. The pro- value, i.e., the flows that do not meet their QoS requirements.
posal addresses both real time (i.e., video conferencing) and Next, the flow with maximum transmission rate is chosen to
BE traffic. The authors formulate the problem of scheduling be scheduled. If none of the flows in the set can transmit
in the OFDMA wireless communications which can be solved due to deep fading, then p is increased and the algorithm
using the dual optimization technique [52]. The proposed builds a new set of eligible users for transmission. The size
algorithm calculates the optimal solution for every frame of the set can be tuned to adjust the trade-off between QoS
which guarantees average throughput with bounded fluctua- provisioning and throughput (i.e., the larger is the set, 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.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 9

higher is the diversity gain). ARTC uses the same approach n 2

as AQDC for scheduling flows from a chosen set. ARTC is 
xi
different from AQDC in the sense that it builds the eligible i
user set based on the residual time obtained for the three Jain’s index = 
n .
adopted QoS objectives. Residual time is the time that a n x2i
i
flow can await before transmission without violating its QoS
requirements. The residual time was originally employed in In [57], authors introduce optimal policies for opportunistic
Urgency Feasibility scheduler [54]. Urgency feasibility sched- scheduling in OFDM systems with three different fairness
uler computes the residual time for all flows and serves the criteria, namely temporal fairness, utilitarian fairness, and
flow with lowest residual time. Furthermore, a call admission minimum-performance guarantees. Under temporal fairness
control policy is proposed to improve the QoS. A flow is criteria, all users are given at least a certain share of airtime,
rejected if the percentage of unsatisfied flows exceeds the where under utilitarian fairness criterion users are given a
admission percentage. In the end, the performance of BQDF, certain share of throughput [58]. The policies with minimum-
AQDC, ARTC, Urgency Feasibility, and MaxRate schedulers performance guarantees, as the name implies, aim to max-
are evaluated via simulation using NS2 [55] under various imize the network performance while satisfying minimum
scenarios. It is shown that AQDC and ARTC perform best user requirements. Temporal and utilitarian fairness methods
in terms of maintaining QoS objectives. In terms of packet oblige the scheduler to allocate a predefined share of re-
drops, AQDC and ARTC have more than 20% less packet drop sources (i.e., time, throughput) to every user. In contrast, with
than others. AQDC and ARTC achieve the same throughput minimum-performance guarantees the scheduler is restrained
as MaxRate which is more than 50% higher than Urgency to satisfy the minimum service requirement of the users.
Feasibility and BQDF. In conclusion, the schedulers that use The authors [57] interpret the optimal policies as bipartite
the newly introduced QoS parameter (i.e., QoS deviation) matching problem and solve them using Hungarian algo-
outperform those which operate on a single QoS objective. rithms [59]. Simulation results shows that temporal, utilitarian,
and minimum-performance guarantee policies provide 46%,
32%, and 31% gain over Round Robin, respectively.
C. Summary One of the most diffused opportunistic approaches with
The authors of [41], [42], [43], [37] propose schedulers fairness constraints is the proportional fair scheduler [7], [60].
with only one QoS metric. Kim et al. propose a scheduler This scheduler assigns priorities to users based on the ratio of
to avoid the performance anomaly caused by traffic integra- two functions: the first function accounts for the rate poten-
tion [41]. The works in [42], [43], [37] are dedicated to delay tially achievable in the current slot, while the second function
improvement. Authors of [42] propose a log rule-based sched- accounts for historical average of the user’s throughput.
uler to achieve delay optimality. Neely uses Lyapunov drift In [61], Tsai introduces four scheduling algorithms that are
and optimization techniques to propose a throughput-optimal based on proportional fair (PF) algorithms. All four proposed
scheduler with worst case delay guarantees and without any schedulers exploit user and channel diversity to improve the
prior knowledge of channel conditions [43]. system performance. PF scheduler prioritizes users based on a
The schedulers in [44], [45] can support multiple QoS balance between the current achievable rate and fairness [62],
metrics. In [44], authors propose a scheduler for real-time [63], [10]. The following illustrates the PF scheduler general
and best effort services using the dual optimization technique. scheduling policy and throughput monitoring [10]:
Authors in [45] propose schedulers which operate based on the
Ri (t)
position of the flow in the QoS unity cube or based on the QoS i∗ (t) = arg max ; (1)
i∈Nb Ti (t)
residual time. Table IV shows each proposal mainly focusing  
on QoS with details regarding the assumptions taken by the 1 Ri (t)
Ti (t + 1) = 1 − Ti (t) + Ii=i∗ (t) ; (2)
authors, analytical tools used for the proposal, the scenario tc tc
in which the proposal in applicable, and other considerations
where i∗ (t) is the user scheduled at time t, Nb is the set
taken into account besides QoS.
of users, Ri (t) and Ti (t) are the peak feasible data rate and
monitored throughput of user i at time slot t, I{.} is the
V. FAIRNESS indicator function and tc is an averaging time window. In
Due to the greedy behavior of opportunistic schedulers, their the PF scheduler, it is assumed that the number of users is
fairness performance is always a concern. Scheduling users fixed (stationary scenario) and queues are fully backlogged.
opportunistically can result in under-serving some users due These assumptions are not realistic and variation in number
to their poor channel quality, while the rest are over-served of users and their queue length has negative impacts on system
because they are in a better channel conditions. As a result, performance. As a remedy, Tsai [61] proposes to stop updating
it is essential to monitor the way a scheduler allocates the Ti (t) for idle users which avoids giving idle users higher
resources to avoid unfairness among users in the long term. priority upon their next transmission. If Ti (t) is updated for
There are different metrics defined for fairness (e.g., Jain’s idle users, upon their activation, the scheduler assumes that
index, temporal fairness, utilitarian fairness). Jain’s index they had packets to transmit within last few slots and they
is one of the popular fairness metrics for studying fair- should be given higher priority to maintain fairness. The
ness performance of the schedulers. For a given set X = four proposed schedulers account for the dynamic changes
{x1 , x2 , . . . , xn } Jain’s index is computed as follows [56]: in number of active users and queue length. These schedulers
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 COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

TABLE IV
S UMMARY OF PROPOSALS WITH MAIN FOCUS ON QoS
Proposal Assumptions Analytical tools Topology Other focus
Flow-level scheduler Non-full CSI Markov decision process Single cell Stability
for wireless networks [37] Not fully backlogged Gilbert-Elliot model Downlink
Traffic: Bernoulli Lagrangian multipliers
Dynamic programming
Scheduler for wireless Not fully backlogged Markov chain Single cell (TDMA) Stability
systems with integrated Traffic: Poisson Lyapunov drift
traffic [41] Foster theorem
Delay-optimal Log rule based Not fully backlogged Markov decision process Single cell (HDR) Throughput-optimal
scheduler for wireless Traffic: Poisson Dynamic programming Downlink Delay-optimal
networks [42] Lyapunov drift
Foster theorem
Opportunistic scheduler for No CSI Lyapunov drift Single/multi-hop
wireless networks with Not fully backlogged Lyapunov optimization
worst-case guarantee [43]
Scheduler for OFDMA Fully backlogged Duality theory Single cell (OFDMA)
systems with multimedia Lagrangian multipliers Downlink
support [44] Convex optimization
Adaptive QoS scheduler Not fully backlogged Unity cube mapping Single cell
for wireless networks [45] Traffic: Poisson & Pareto Downlink

employ different methods with different level of sophistication not affect the QoS. On the contrary, satisfaction of services
to update users’ Ti (t). The first two algorithms only update such as video streaming depends on the instantaneous data
Ti (t) for active users and their difference is only the initial rate. For such services, the utility function should be based on
value of Ti (t). The third algorithm uses a more sophisticated instantaneous rate. Kwon et al. propose a framework that can
approach by updating Ti (t) according to the status of the user accommodate both elastic and non-elastic services. Instead of
in the previous slot (i.e. active or idle) and the last time it is deterministic scheduling, they use a probabilistic scheduling
scheduled. Finally, the fourth algorithm changes the averaging policy that randomly schedules a user per time-slot with a
window size tc with reference to the backlog of the users. certain probability. Kwon et al. model the channel using a
Results show that proposed algorithms reduce the impact of finite state Markov chain (FSMC), formulate the scheduling
dynamic changes in the network on packet delay and data rate. problem based on convex optimization [67] and solve it using
The authors of [64] adapt the analytical model proposed a Lagrangian function, the duality theorem [68]. An iterative
for PF by Liu et al. [7] to an OFDMA networks with algorithm is also proposed that can find the optimum solution
more realistic assumptions. Their model accounts for multiple in every time slot. It is shown via numerical simulation that
subcarriers, but also for less realistic Poisson traffic arrivals. their proposed scheduler meets the required fairness objective
The adapted PF scheduler computes a matrix containing user for users with elastic and non-elastic services.
rankings over all subcarriers. For every subcarrier, the user
with the highest rank and non-empty buffer is scheduled. Non- A. Summary
empty buffer condition accounts for the fact that in real world a
users can be eligible to be scheduled when it has no packets to The surveyed literature in this section considered fairness in
transmit. In such cases, the scheduler selects the next best user opportunistic schedulers. As pointed out in [57], different met-
with non-empty buffer to avoid wasting airtime. The proposed rics can be used to evaluate the fairness in a transmission sys-
analytical model is validated in terms of average throughput tem, e.g., temporal fairness, utilitarian fairness, and minimum-
and Jain’s fairness index by simulation. performance guarantees. Proportional fair schedulers attempt
to balance between users with the best channel state, and
In [65], an adaptive resource allocation for OFDM system
users which received less throughput in the past [7]. Tsai,
is proposed that accounts for each user’s required data rate
in [61], and Almatarneh et al., in [64], point out the importance
as a fairness measure. The authors formulate the optimization
of throughput monitoring with proportional fair schedulers.
problem for subchannel and power allocation with a propor-
Tsai proposes four variants of the original proportional fair
tional fairness constraint. Since their proposed optimization
algorithm that accounts for parameters such as users activity
problem requires linearization of nonlinear constraints, the
log and backlog. Kwon et al. [66] propose a fair opportunistic
authors propose a suboptimal solution with lower complexity.
scheduler based on probabilistic scheduling approach using
The suboptimal solution carries out the subchannel allocation
convex optimization and dual theory. Table V shows each
and power allocation separately. Via simulations, it is shown
proposal mainly focusing on fairness with details regarding
that the suboptimal solution can achieve 95% of the optimal
the assumptions taken by the authors, analytical tools used for
capacity with much lower complexity.
the proposal, the scenario in which the proposal in applicable,
In [66], Kwon et al. tackle the fairness issue in TDMA
and other considerations taken into account besides fairness.
networks. The proposed opportunistic fair scheduler prior
to [66] are either based on average rate-based utility functions
or instantaneous rate-based utility functions. Average rate- VI. D ISTRIBUTED S CHEDULING A LGORITHMS
based utility functions are suitable for elastic services (e.g., In a centralized scheduling approach, the scheduler is aware
HTTP, Email, and FTP) for which instantaneous data rate does of all user’s channel condition or it will acquire an estimate
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 11

TABLE V
S UMMARY OF PROPOSALS WITH MAIN FOCUS ON fairness
Proposal Assumptions Analytical tools Topology Other focus
Proportional fair scheduler Full CSI Single cell (TDM)
for wireless forward link Not fully backlogged Downlink
data services [61] Traffic: Poisson
Proportional fair scheduler Full CSI Convex optimization Single cell (TDMA) Jain’s fairness Index
for wireless forward link Not fully backlogged Duality theory
data services [66] Traffic: Poisson Lagrangian multipliers
Analytical model for Not fully backlogged Single cell (OFDMA) Jain’s fairness Index
proportional fair scheduler
in OFDMA systems [64]
Opportunistic schedulers Full CSI Maximal bipartite matching Single cell (OFDM) Temporal fairness
optimized under different Fully backlogged Hungarian algorithms Downlink Utilitarian fairness
fairness consideration [57] Minimum performance guarantee
Adaptive opportunistic scheduler Full CSI Linear programming Single cell (OFDM)
for Multiuser OFDM Non-linear programming
Networks [65] Newton-Raphson method [69]

rate threshold based on the local information of each user.

Next, they show that it is optimal to transmit when the current
transmission rate is higher than an optimal threshold.
The second scenario analyzed in [70] is a non-cooperative
game with selfish users trying to increase their throughput.
Zheng et al. use game theory to model the behavior of
Fig. 6. DOS transmission example. users in this network. The authors investigate the existence
and uniqueness of the Nash equilibrium [72] using the best
response strategy. For both scenarios (i.e., cooperative and
non-cooperative), the optimal rate can be computed via an
of that information to make the scheduling decision. On the iterative algorithm based on the local information. Using local
other hand, in a distributed scheduling approach, users make information alleviates the need for message exchange between
scheduling decisions independent of the central entity and users for reporting their CSI. Numerical simulations indi-
possibly without an overall knowledge of the network. cates that applying optimal stopping theory results in higher
Distributed opportunistic scheduling (DOS) did not get throughput gains. It is also shown that the non-cooperative
much attention until recent years. The authors of [70] and [71] scheme is less efficient than the cooperative scheme due to
took the first steps to study such systems under various selfish behavior of users. However, the inefficacy of the non-
scenarios. In [70], Zheng et al. study DOS in ad-hoc networks cooperative scheme can be mitigated by forcing a price-based
with random channel access under two scenarios: 1) all users mechanism.
cooperate to maximize the aggregate network throughput The DOS proposed in [71] is an extension to the work
(i.e., cooperative); and 2) each user tries to maximize its in [70] with the addition of average delay constraint. Hence,
own throughput (i.e., non-cooperative). In the first scenario the proposed algorithm should maximize the throughput for
(cooperative), all users contend for the channel but successful both cooperative and non-cooperative scenarios mentioned
contention is not necessarily followed by transmission. Upon in [70], while maintaining the designated delay constraints.
successful contention, the user transmits only if it is in good The authors consider both network-wide and per-user average
channel condition (i.e., high data rate), otherwise it allows the delay constraints. As for the network-wide average delay, they
rest of the users to re-contend for the channel. This process seek for the optimal transmission rate value in which all
goes on until a user in good channel condition wins the users cooperate to increase the total network capacity while
contention. The cooperation benefits all users because it leads maintaining the average delay less than α. In the case of
to network capacity increment in the long run. Like any other individual user average delay constraint, that is defined for
opportunistic approach, the cooperation gain is depending on the non-cooperative scenario, each user tries to maximize its
the channel diversity of the users. It can be seen that there is a throughput while keeping its average delay less than αm . In
trade-off between the gain obtained from probing more users order to find the optimum threshold, Tan et al. again employ
and the time it takes to find a user with high rate. In Fig. 6, optimal stopping theory to formulate the problem and solve it
we can see that total available time Ttotal is divided between using a stochastic Lagrangian approach. The results show that
n contention period Tn and data transmission period Ttran . the optimal transmission rate threshold is upper bounded by a
Clearly, the channel transmission time is reduced when function of α if the average delay is less than a critical time
more users probe the channel. Hence, there is a threshold be- (α < α∗ ). In other words, the average delay constraint does
yond which, re-contending for channel becomes sub-optimal. not affect the optimal threshold policy if (α > α∗ ); otherwise
The question is: what is the optimal transmission rate so that a the optimal threshold policy will be a function of α. This result
user does not need to drop its transmission chance and allow also applies to the non-cooperative scenario, where there exists
other users to re-contend for channel? The authors propose an α∗i for every user i. The authors evaluate the performance of
an online iterative algorithm that can compute the optimal the proposed DOS using numerical simulations which indicate
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 COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

users are also affected by the interference from neighboring

cells, see Fig. 7. Indeed, it has been shown that Inter-Cell
interference (ICI) can be reduced by exchanging information
between base stations using multi-cell scheduling [75], [76],
[77], [78], [79], [80], [81], [82].
In the multi-cell scheduling approach, every base station
first computes the ICI from the signal measurements sent by its
users, and then exchanges this information with neighboring
base stations. A drawback of this approach is that a high
speed connection should be allocated to inter-base station
communications. Another drawback is the high computation
overhead on the central entity that processes this information
for further scheduling decisions (EPC in LTE [2]).
Fig. 7. Inter-Cell interference in cellular networks. Tang et al. [75] propose two joint multi-cell scheduling and
beam coordination schemes, namely SINR feedback and ABC.
In the SINR feedback scheme, each base station selects m
that higher values of SNR reduce the effect of α on the optimal beams randomly and sends beam-pilots within the cell. The
policy. users will send their feedback with respect to their SINR
The authors of [73] propose a two-level probing for the which are also affected by the beam pilots received from
DOS proposed in [70]. In practice, the channel estimation the neighboring cells. Each base station makes scheduling
obtained from the first time probing is not very accurate due decisions based on the received SINR feedbacks. Intra-cell
to noise. Hence, the authors [73] propose a two-level probing scheduling decisions are made based on the PF algorithm
policy which appears to be a threshold-based policy. If the proposed in [83].
first channel estimation falls in between thresholds, then it is In the ABC scheme, the cellular network is partitioned di-
optimal to probe the channel for the second time. It is shown viding base stations into A/B/C subsets. Although the network
via simulation that the two-level probing policy can increase is partitioned, it operates as a reuse factor 1. In the first step,
the gain over one-level probing policy up to 110%. base stations tagged as A select m beams, send beam pilots
Garcia-Saaverdra et al., in [74], address the shortcomings and make their scheduling decision. Note that users of subset
of previous DOS proposals such as [71], [70], [73] and pro- B and C also listen to beam pilots to be able to estimate
pose an adaptive distributed opportunistic scheduling (ADOS). their SINR. After the scheduling decision has been made, the
Previous DOS proposals were designed for a scenario with identity of the scheduled users and the beams assigned to them
fully saturated users (i.e., fully backlogged user’ transmission is sent to the neighboring base stations tagged as subset B.
queues). In a real networks users have different queue sizes, This helps the base stations in subset B to avoid choosing
in fact there are situations when user contends for the channel beams that interfere with subset A. Base station in subset B
but it has not enough data to transmit during the whole will perform the same operation and inform to those in subset
contention period. Hence, there are periods of time in which C. Due to priority (in term of frequency selection) given to
no one is transmitting. These non-transmission periods are the base stations in subset A, they will provide better service
called empty mini-slots in [74]. This is the drawback of DOS to their users in comparison to the base stations in subset B
which makes it sub-optimal for unsaturated scenarios. ADOS and C. To avoid such unfairness, the authors propose to assign
overcome the aforementioned drawback by controlling both the A/B/C tags in a Round Robin fashion.
the optimal transmission rate threshold (i.e., the rate at which In their work, Tang et al. compare the proposed schemes
users transmit after a successful contention) and the access with two other schemes. In the first scheme, network operates
probability (i.e., the probability that a user attempts to contend using a frequency reuse partition with a factor 3. In the second
for the channel). The authors exploit access probability to dis- scheme, each base station operates fully independently and
criminate among users with large backlog and those with small without considering ICI. Using simulations, it is shown that
backlog. Hence, the probability of empty mini-slot occurrence SINR feedback and ABC outperform the two other schemes.
is reduced by assigning higher access probabilities to users ABC scheme provides more than 100% throughput gain in
with large backlog. The optimal rate and access probabilities comparison to schemes without ICI control mechanism. In ad-
are computed locally via a simple algorithm based on control dition, users on the edge of the cell receive higher throughput
theory that adapts itself to the current state of the network with ABC scheme.
and converges to the optimal operation point. The drawback of In [76], Bendlin et al. propose a distributed multi-cell
ADOS can be the convergence time of the adaptive algorithm. scheduling, namely cooperative eigen beamforming (CEB),
Using simulations, the authors show that ADOS outperforms that is tolerant to delay and capacity limitations of backhaul
previous DOS proposals in terms of proportional fairness, and links. In many proposals, authors assume that backhaul links
achieves 30% higher throughput. have zero delay and unlimited capacity, which is not a realistic
All previously reviewed distributed approaches suit carrier assumption. In CEB, authors assume that each base station
sense multiple access (CSMA). However, distributed schedul- schedules one user per slot and it has full CSI knowledge of
ing algorithms can also be used in multi-cell OFDMA schedul- its users. Scheduling is performed in two steps in CEB. In the
ing. In a real network, apart from environmental noise, mobile first step, each base station chooses the proper beamforming
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 13

TABLE VI
S UMMARY OF PROPOSALS WITH distributed scheduling
Proposal Assumptions Analytical tools Topology Other focus
Distributed opportunistic scheduler Non-full CSI Optimal stopping theory Single hop (ad-hoc) Throughput
for ad-hoc networks [70] Fully backlogged Nash equilibrium Random access Fairness
Game theory
Distributed opportunistic scheduler Non-full CSI Optimal stopping theory Single hop (ad-hoc) Throughput
for ad-hoc networks under Fully backlogged Duality theory Random access Fairness
delay constraints [71] Game theory Nash equilibrium
Stochastic Lagrangian approach
Distributed opportunistic scheduler Non-full CSI Optimal stopping theory Single hop (ad-hoc) Throughput
for ad-hoc networks with Fully backlogged Lebseque’s convergence theorem [84] Random access
two-level probing [73]
Distributed opportunistic scheduler Non-full CSI Control theory Single hop (ad-hoc) Throughput
for ad-hoc networks using Not fully backlogged Random access Fairness
control theory [74] Traffic: Stations transmitting Stability
at half saturation rate
Multi-cell opportunistic scheduler Non-full CSI Multi-cell Throughput
for cellular networks using Downlink Fairness
random beamforming [75]
Multi-cell scheduling with Full CSI Multi-cell Throughput
consideration of delay and capacity Error-free backhaul Downlink Fairness
limitations in backhaul links [76]

TABLE VII
which minimizes ICI and in the second step, a user will be C OMMON TECHNIQUES USED IN OPPORTUNISTIC SCHEDULING
scheduled. The main advantage of CEB is the low amount Category Analytical tools
of data exchanged among base stations. Bendlin et al. show - Markov chain [28], [31], [20], [38], [21], [23], [37]
that CEB can perform very close to schemes that disseminate - Lyapunov [28], [19], [20], [34], [21], [22]
- Dynamic programming [19], [23], [37]
the full CSI information in the network. CEB also exhibits Capacity - Optimal stopping theory [34]
robustness towards delay in backhaul links. - Greedy algorithms [19]
- Restless multi-armed bandit [37], [31]
- Lagrangian multiplier [28], [19], [20], [37]
- Markov chain [41], [42]
A. Summary - Lyapunov [41], [42], [43]
In this section we have reviewed proposals focusing on QoS - Dynamic programming [42]
- Duality theory [44]
distributed opportunistic scheduling strategies. The body of - Lagrangian multiplier [44]
work [70], [71], [74], [73] investigates distributed scheduling - Markov chain [66]
for wireless networks with random medium access (e.g., Fairness - Duality theory [66]
- Hungarian algorithms [57]
CSMA). In [70], [71] a distributed scheduler is proposed, - Optimal stopping theory [71], [70], [74], [73]
which uses optimal stopping theory to make transmission de- Distributed - Nash equilibrium [70], [74]
cisions according to the channel state. The scheduler proposed - Control theory [74]
- Duality theory [71]
in [71] includes delay constraints in addition to throughput. - Game theory [71], [70]
In [74] a distributed algorithm based on control theory is
proposed that improves the proposals in [71], [70], [73] by
considering dynamic traffic conditions.
Proposals in [75], [76] focus on distributed scheduling in
cellular networks. Tang et al. [75] propose two approaches, A. Common techniques
namely SINR feedback and ABC, to reduce the ICI in a
multi-cell scenario. Bendlin et al. [76] use cooperative eigen
beamforming for ICI reduction. Table VI shows each proposal Markov chain is commonly used in many papers, especially
focusing on distributed scheduling with details regarding the for modeling the channel variations in wireless networks [41],
assumptions taken by the authors, analytical tools used for the [42], [28], [31], [85], [20], [34], [21], [23], [66], [37]. The
proposal, the scenario in which the proposal in applicable, and Lyapunov drift and optimization techniques are widely em-
other considerations taken into account. ployed to prove the stability of the schedulers or to opti-
mize them [19], [20], [28], [34], [21], [22], [41], [42], [43].
VII. S URVEY ’ S SUMMARY Many proposals exploit other mathematical techniques such
as duality theory [44], [66], optimal stopping theory [34],
In this section, we conclude the work with summarizing the [71], [70], [74] and restless multi-armed bandit [31], [37] for
common techniques, evaluation methods, and possible future optimization purposes. It was of surprise that the common
directions of research in opportunistic scheduling algorithms. optimization techniques (e.g., linear programming, non-linear
Moreover, the proposals under different categories are summa- programming) are rarely used in recent works. Table VII
rized in Table VIII. We also provide a performance comparison illustrates the popularity of each technique in every category
among different schemes based on the results provided in the of opportunistic scheduling, according to the taxonomy we
reviewed papers. defined in Section II.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
14 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

TABLE VIII
P ERFORMANCE COMPARISON OF DIFFERENT CATEGORIES OF OPPORTUNISTIC SCHEDULING ALGORITHMS .
Category Assumptions Analytical tools Topology Other focus Achieved improvement
-Non-full CSI -Markov chain -Single cell -Throughput -Up to 37% throughput
-Full CSI -Lyapunov drift -Multi-cell -Throughput-optimal gain over Round Robin.
-Not fully backlogged -Lyapunov optimization -OFDMA -Near throughput-optimal -Up to 70% delay reduction
-Fully backlogged -Lagrangian multipliers -OFDM -Delay compared to Round Robin.
-Traffic: Generic, -Finite state Markov chain -HSDPA -Fairness -Up to 95% average queue size
Poisson, Bernoulli -Markov decision process -FDD -Stability improvement over MaxWeight.
Capacity -Error free transmission -Partially observable -Downlink and -Up to 4% improvement over
Markov decision process generic expected utility compared to
-Dynamic programming optimal policy.
-Linear programming
-Restless multi-armed
bandit process
-Optimal stopping theory
-Gilbert-Elliot model
-Maximum weight match
-Non-full CSI -Markov chain -Single cell -Throughput -Between 30∼40% throughput
-Not fully backlogged -Lyapunov drift -Single hop -Throughput-optimal gain over M-LWDF.
-Fully backlogged -Lyapunov optimization -multi-hop -Delay-optimal -Achieving up to 92% of
-Traffic: Poisson, -Lagrangian multipliers -TDMA -Delay maximum achievable throughput.
Pareto -Foster theorem -OFDMA -Stability -Average delay gain of 50%
QoS
-Convex optimization -HDR over MaxRate.
-Dynamic programming -Downlink and -Between 20∼80% delay
-Duality theory generic improvement over Exp rule
-Unity cube mapping which itself outperform
-Maximum weight match MaxWeight.
-Full CSI -Convex optimization -Single cell -Jain’s fairness -Achieves a minimum Jain’s
-Not fully backlogged -Lagrangian multipliers -OFDMA -Temporal fairness index of 0.6 with varying
-Fully backlogged -Duality theory -OFDM -Utilitarian fairness data rate.
Fairness -Traffic: Poisson -Maximal bipartite matching -TDMA -Minimum performance -Up to 46% throughput
-Hungarian algorithms -TDM guarantee improvement over Round Robin.
-Linear programming -Downlink and -Delay improvement of 12%
-Non-linear programming generic over PFA.
-Newton-Raphson method
-Full CSI -Optimal stopping theory -Single hop -Throughput -Throughput gain of 100% over
-Not-full CSI -Nash equilibrium -Multi-cell -Fairness an independent scenario.
-Fully backlogged -Duality theory -Random access -Stability -Up to 40% throughput gain
-Not fully backlogged -Stochastic Lagrangian -Ad-hoc over the DOS proposed in [70]
Distributed -Traffic: Stations approach -Downlink and that obtains 13.9∼42.8% gain
transmitting at half -Lebseque’s convergence generic over the proposal in [86] that
saturation rate theorem improves the throughput of
-Error free backhaul -Control theory existing protocols up to 50%.
communications -Game theory

B. Evaluation Method limited computational capacity of base stations and its affect
The proposed opportunistic algorithms can be evaluated via on scalability of the network. The following is a list of most
numerical calculation obtained from analysis and simulations common assumptions:
using available network simulators (e.g., NS2 [55], NS3 [87], • Full CSI availability of mobile users’ channel at base

OPNET [88], etc.). The nature of opportunistic schedulers station.

requires cross-layer information that is not always available • Fully or infinitely backlogged queues.

at MAC layer. The time and effort required to adapt a new al- • Mobile users with i.i.d. channel distributions.

gorithm into the current protocols makes implementation less • Fixed number of mobile users (no one leaves or joins the

popular than numerical results. Hence, most of the proposals channel).

are supported by mathematical analysis and simulation via • Scheduling decision can be made offline.

ad-hoc simulators. As such, the evaluation does not consider • Scheduling single user per frame.

the impact of the proposed schedulers on other layers (e.g., • Scheduling in single channel network.

application, network, and transport) and vice versa. Some of these assumptions can highly affect the perfor-
mance of a proposal. For instance, base station obtains the
CSI via feedback received from users. Assuming that every
C. Open issues user is sending feedbacks on every slot and for every available
So far we have reviewed opportunistic scheduling from subchannel, a big chunk of bandwidth is wasted for feedbacks.
various aspects. In this subsection, we highlight the open Another example is the assumption of fully backlogged queues
research issues in opportunistic scheduling. that made the MaxWeight [13] scheduler optimal, although, it
1) Optimality under realistic assumptions: There are many was proven to be sub-optimal under queue dynamics (i.e., not
proposals in opportunistic scheduling that are optimal under fully backlogged queues).
certain assumptions. Authors of [17], [89] focus on this fact in There have been efforts to eliminate such unrealistic as-
OFDMA systems by showing the real world constraint such as sumptions. However, by removing these assumptions, we
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 15

would increase computation and modeling complexity. We cooperate for detecting more accurate spectrum opportuni-
believe that opportunistic scheduler proposals are not imple- ties [99]. For instance, in a very recent work, the authors
mented in practice due to high complexity of the proposals of [100] propose a cooperative packet delivery for hybrid
and their unrealistic assumptions which may results in scarce wireless networks. They use coalitional game theory to group
interest for practical applications. It would be interesting to mobile users that assist each other by relaying traffic. Today’s
investigate the performance of optimal schedulers in real cellular phones are equipped with a secondary interface (e.g.
implementations or under realistic assumptions. WiFi) as well as the cellular interface. The authors propose
2) Fairness: Fairness is also an open issue in opportunistic to take advantage of this capability to form coalitions among
schedulers. In comparison with other categories of schedulers, mobile users using the secondary interface. Once a coalition is
fairness in opportunistic scheduling is an under-explored field. formed, coalition members can send their traffic, on secondary
In a real network, it is crucial for the operator to achieve the interface, to the user in the best channel to relay their traffic
service rate committed to users. A first step toward encourag- to cellular network.
ing operators to implement opportunistic schedulers into their 5) Implementation: To the best of our knowledge, no one
networks is a proposal with tight fairness control features. explored the gain obtained via opportunism in an actual
As mentioned previously, there is always trade-off between experiment. Hence, we believe such implementations would
fairness and opportunism and the challenge is to decide when provide insights into unforeseen problems and consequences
a scheduler should stop being opportunistic and starts being of opportunistic scheduling.
fair. However, increasing fairness does not always result in
throughput reduction. For example when multiple users are
experiencing the same channel conditions, the scheduler can D. Future trends
improve the fairness without affecting the throughput. In other After two decades of research, we can say that opportunistic
words, in the event that multiple users can transmit at the scheduling became very mature. This maturity calls for new
same rate, the scheduler can tune the fairness by allowing the opportunistic schedulers whose designs are backed up with
least served user to transmit. As a result, the fairness can be analysis. Indeed, many authors not only show the performance
improved without throughput impairment. advantage of their proposal, but also prove the stability of the
3) Data offloading: Nowadays, many mobile devices are schedulers.
equipped with WiFi. In [90] and [91], the authors propose to Currently, researchers are active towards two major direc-
take advantage of WiFi to offload the elastic traffic on WiFi tions. First, researchers evaluate the performance of existing
and real time traffic on cellular network. Since a large portion proposals under more realistic scenarios such as flow-level
of cellular traffic is formed by elastic traffic, opportunistic dynamics, multi-user multi-carrier scheduling, and mobility.
data offloading can reduce the congestion in cellular networks This helps us to have a better overview of the system per-
and improve the QoS of users. Data offloading can also be formance in a real world scenario. Second, researchers seek
performed by exploiting the user’s behavior. In event that for novel applications and new challenges for opportunistic
people with shared interest gather in one place (e.g., an sport scheduling. Use of opportunistic scheduling in cooperative
match), it is very likely that they try to access the same content communications is one of the newly explored areas which
multiple times. In such case, users who already acquired the has attracted the interest of many researchers.
content, can offload the content to users who are requesting An interesting example to integrate opportunistic scheduling
for it. This approach benefits cellular network by saving with cooperative communications would be forming clusters
bandwidth for re-sending the same content. In addition, it can among mobile users. In this scenario, each cluster chooses the
save the network from saturation in places where the density member experiencing the best channel to handle the cluster
of networks varies much (e.g. stadium). traffic.
4) Cooperative networks: We believe that opportunistic
scheduling can effectively take advantage of cooperative di-
VIII. C ONCLUSIONS
versities created in cooperative networks [92], [93], [94]. For
example, users with good channel quality can forward data for In this paper, we categorize the opportunistic scheduling
the users with poor channel quality. Hence, using, e.g., relay proposals based on the approach they took for formulating the
nodes among the mobile users, the scheduler would be not problem of scheduling, i.e., capacity, QoS, fairness, distributed
required to wait for channel quality improvements of users in scheduling.
bad channel states, or waste bandwidth by scheduling users The opportunistic proposals that aim to improve the system
with low transmission rates. As observed in [22], users can capacity are able to achieve throughput optimality. However,
benefit from cooperation both individually and network wise. these proposals may not be able to guarantee any fairness
There are several approaches for cooperative communications, and QoS metrics. QoS related opportunistic schedulers can
such as cooperative MIMO [95], [96], [97], and relaying [98]. also achieve throughput-optimal or near throughput-optimal
In cooperative MIMO or virtual antenna array, users with results while maintaining the QoS constraints. Throughput
single antenna join and create a virtual MIMO device [97]. gain obtained by these proposals depends on the number
There are various relaying techniques but they all operate of QoS metrics and their desired values. We also observed
based on the concept of forwarding data to an entity with better that the unfairness issue which is caused by the greedy
or more reliable channel [98]. Another interesting scheme nature of opportunistic schedulers can be resolved at the cost
is cooperative sensing in cognitive network in which users of throughput reduction. Although these schedulers achieve
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
16 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

fairness at the cost of throughput reduction, they still out- [17] L. Badia, A. Baiocchi, A. Todini, S. Merlin, S. Pupolin, A. Zanella, and
perform the non-opportunistic schedulers. In order to reduce M. Zorzi, “On the impact of physical layer awareness on scheduling
and resource allocation in broadband multicellular IEEE 802.16 sys-
the complexity of opportunistic schedulers, some proposed tems [radio resource management and protocol engineering for IEEE
opportunistic distributed scheduling. 802.16],” IEEE Wireless Commun., vol. 14, no. 1, pp. 36–43, 2007.
Opportunistic scheduling appears to be a promising solution [18] A. Asadi and S. W. Tan, “An enhanced cross layer downlink scheduling
algorithm for IEEE 802.16 networks,” in Proc. IEEE ICOIN, January
to wireless channel bandwidth limitations. In this survey, the 2011, pp. 212 –217.
importance of opportunistic schedulers in wireless networks [19] M. Andrews and L. Zhang, “Scheduling algorithms for multi-carrier
was emphasized with extensive reviews and comparisons wireless data systems,” in Proc. ACM MobiCom. ACM, 2007, pp. 3
–14.
among them. It was observed that the proposals with focus on
[20] S. Liu, L. Ying, and R. Srikant, “Scheduling in multichannel wireless
capacity can provide up to 37% capacity gain and 70% delay networks with flow-level dynamics,” in Proc. the ACM SIGMETRICS,
reduction compared to Round Robin. QoS oriented proposals 2010, pp. 191–202.
offered throughput improvement up to 30% throughput gain [21] C. Luo, F. R. Yu, H. Ji, and V. C. Leung, “Optimal channel access for
TCP performance improvement in cognitive radio networks,” Wireless
and delay improvement up to 80%. Proportional fair based Networks, vol. 17, pp. 479–492, February 2011.
schedulers achieved a minimum Jain’s index of 0.6 while [22] K. Khalil, M. Karaca, O. Ercetin, and E. Ekici, “Optimal scheduling in
maintaining the QoS requirements. Finally distributed schedul- cooperate-to-join cognitive radio networks,” in Proc. IEEE INFOCOM,
April 2011, pp. 3002 –3010.
ing proposals appeared to have high throughput gain without [23] H. Al-Zubaidy, I. Lambadaris, and J. Talim, “Optimal scheduling in
the need for a central managing entity. We also introduced high-speed downlink packet access networks,” ACM Trans. Modeling
some of the open issues and discussed possible solutions based and Computer Simulation, vol. 21, pp. 3:1–3:27, December 2010.
on the state-of-the-art proposals and researches done within [24] L. Tassiulas, “Scheduling and performance limits of networks with
constantly changing topology,” IEEE Trans. Inf. Theory, vol. 43, no. 3,
past two decades. pp. 1067 –1073, May 1997.
[25] M. Neely, E. Modiano, and C. Rohrs, “Dynamic power allocation
and routing for time-varying wireless networks,” IEEE J. Sel. Areas
R EFERENCES Commun., vol. 23, no. 1, pp. 89 – 103, January 2005.
[26] A. Eryilmaz and R. Srikant, “Fair resource allocation in wireless
[1] 3GPP, “Dual-cell high speed downlink packet access (HSDPA) opera- networks using queue-length-based scheduling and congestion control,”
tion (release 8),” 3GPP TR 25.825. in Proc. INFOCOM, vol. 3, March 2005, pp. 1794 – 1803 vol. 3.
[2] H. Ekstrom, A. Furuskar, J. Karlsson, M. Meyer, S. Parkvall, J. Torsner, [27] A. Lyapunov, “The general problem of the stability of motion,”
and M. Wahlqvist, “Technical solutions for the 3G long-term evolu- International J. Control, vol. 55, no. 3, pp. 531–534, 1992.
tion,” IEEE Commun. Mag., vol. 44, no. 3, pp. 38 – 45, March 2006. [28] S. Liu, L. Ying, and R. Srikant, “Throughput-optimal opportunistic
[3] IEEE 802.16-2004, “IEEE standard for local and metropolitan area scheduling in the presence of flow-level dynamics,” IEEE/ACM Trans.
networks part 16: Air interface for broadband wireless access systems Netw., vol. 19, no. 4, pp. 1057 –1070, August 2011.
amendment 3: Advanced air interface,” 2004. [29] L. Kleinrock, “Queueing systems. volume 1: Theory,” 1975.
[4] I. Mansfield. (2012, March) Mobile internet usage is doubling year on [30] D. Bertsekas, D. Bertsekas, D. Bertsekas, and D. Bertsekas, “Dynamic
year. [Online]. Available: http://www.cellular-news.com/story/53388. programming and optimal control,” 1995.
php [31] W. Ouyang, S. Murugesan, A. Eryilmaz, and N. Shroff, “Exploiting
[5] , (2012, March) Wireless network data problems increase channel memory for joint estimation and scheduling in downlink
as more subscribers use web apps. [Online]. Available: http: networks,” in Proc. IEEE INFOCOM, April 2011, pp. 3056 –3064.
//www.cellular-news.com/story/53329.php
[32] K. Glazebrook, H. Mitchell, and P. Ansell, “Index policies for the
[6] M. L. Pinedo, Scheduling: Theory, Algorithms, and Systems, 4th ed. maintenance of a collection of machines by a set of repairmen,”
Springer, 2012. European Journal of Operational Research, vol. 165, no. 1, pp. 267 –
[7] E. Liu and K. Leung, “Proportional fair scheduling: Analytical insight 284, 2005.
under rayleigh fading environment,” in Proc. IEEE WCNC, 2008, pp.
[33] S. Guha, K. Munagala, and P. Shi, “Approximation algorithms for
1883–1888.
restless bandit problems,” J. ACM (JACM), vol. 58, pp. 3:1–3:50, 2010.
[8] R. Knopp and P. Humblet, “Information capacity and power control
[34] P. Chaporkar, A. Proutiere, H. Asnani, and A. Karandikar, “Scheduling
in single-cell multiuser communications,” in Proc. IEEE ICC, vol. 1,
with limited information in wireless systems,” in Proc. ACM MobiHoc,
June 1995, pp. 331 –335 vol.1.
2009, pp. 75–84.
[9] , “Information capacity and power control in single-cell multiuser
communications,” in Proc. IEEE ICC, vol. 1. IEEE, 1995, pp. 331– [35] M. Neely, “Stochastic network optimization with application to com-
335. munication and queueing systems,” Synthesis Lectures on Communi-
cation Networks, vol. 3, no. 1, pp. 1–211, 2010.
[10] P. Viswanath, D. Tse, and R. Laroia, “Opportunistic beamforming using
dumb antennas,” in IEEE Trans. Inf. Theory, June 2002, p. 449. [36] G. Peškir, A. Shiryaev, and A. Shiriaev, Optimal stopping and free-
[11] N. Sharma and L. Ozarow, “A study of opportunism for multiple- boundary problems. Birkhauser, 2006, vol. 10.
antenna systems,” IEEE Trans. Inf. Theory, vol. 51, no. 5, pp. 1804 – [37] P. Jacko, “Value of information in optimal flow-level scheduling of
1814, May 2005. users with Markovian time-varying channels,” J. Performance Evalu-
[12] X. Qin and R. Berry, “Exploiting multiuser diversity for medium access ation, vol. 68, no. 11, pp. 1022–1036, November 2011.
control in wireless networks,” in Proc. IEEE INFOCOM, vol. 2. IEEE, [38] M. Ouyang and L. Ying, “On optimal feedback allocation in multi-
2003, pp. 1084–1094. channel wireless downlinks,” in Proc. ACM MobiHoc. ACM, 2010,
[13] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, R. Vijayakumar, pp. 241–250.
and P. Whiting, “Scheduling in a queuing system with asynchronously [39] X. Lin and N. Shroff, “The impact of imperfect scheduling on cross-
varying service rates,” Probability in the Engineering and Informa- layer rate control in wireless networks,” in Proc. IEEE INFOCOM,
tional Sciences, vol. 18, pp. 191–217, April 2004. vol. 3, March 2005, pp. 1804 – 1814 vol. 3.
[14] P. van de Ven, S. Borst, and S. Shneer, “Instability of maxweight [40] P. Chaporkar, K. Kar, and S. Sarkar, “Throughput guarantees through
scheduling algorithms,” in Proc. IEEE INFOCOM, April 2009, pp. maximal scheduling in wireless networks,” in Proc. 43d Annual Aller-
1701 –1709. ton Conference on Communication, Control and Computing, 2005, pp.
[15] J.-W. Lee, R. Mazumdar, and N. Shroff, “Opportunistic power schedul- 28–30.
ing for multi-server wireless systems with minimum performance [41] H. Kim and G. de Veciana, “Losing opportunism: Evaluating service
constraints,” in Proc. IEEE INFOCOM, vol. 2, March 2004, pp. 1067 integration in an opportunistic wireless system,” in Proc. IEEE INFO-
– 1077 vol.2. COM, May 2007, pp. 982 –990.
[16] C. Wong, R. Cheng, K. Lataief, and R. Murch, “Multiuser OFDM [42] B. Sadiq, S. J. Baek, and G. de Veciana, “Delay-optimal opportunistic
with adaptive subcarrier, bit, and power allocation,” IEEE J. Sel. Areas scheduling and approximations: The log rule,” in Proc. IEEE INFO-
Commun., vol. 17, no. 10, pp. 1747–1758, 1999. COM, April 2009, pp. 1692 –1700.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ASADI and MANCUSO: A SURVEY ON OPPORTUNISTIC SCHEDULING IN WIRELESS COMMUNICATIONS 17

[43] M. Neely, “Opportunistic scheduling with worst case delay guarantees [70] D. Zheng, W. Ge, and J. Zhang, “Distributed opportunistic scheduling
in single and multi-hop networks,” in Proc. IEEE INFOCOM, April for ad hoc networks with random access: An optimal stopping ap-
2011, pp. 1728 –1736. proach,” IEEE Trans. Inf. Theory, vol. 55, no. 1, pp. 205 –222, January
[44] K. W. Choi, W. S. Jeon, and D. G. Jeong, “Resource allocation 2009.
in OFDMA wireless communications systems supporting multimedia [71] S.-S. Tan, D. Zheng, J. Zhang, and J. Zeidler, “Distributed opportunistic
services,” IEEE/ACM Transactions on Networking, vol. 17, no. 3, pp. scheduling for ad-hoc communications under delay constraints,” in
926 –935, June 2009. Proc. IEEE INFOCOM, March 2010, pp. 1 –9.
[45] R. K. Beidokhti, M. H. Yaghmaee Moghaddam, and J. Chitizadeh, [72] M. Osborne and A. Rubinstein, A course in game theory. The MIT
“Adaptive QoS scheduling in wireless cellular networks,” Wireless press, 1994.
Networks, vol. 17, pp. 701–716, April 2011. [73] P. Thejaswi, J. Zhang, M. Pun, H. Poor, and D. Zheng, “Distributed
[46] S. Patil and G. de Veciana, “Managing resources and quality of service opportunistic scheduling with two-level probing,” IEEE/ACM Trans.
in heterogeneous wireless systems exploiting opportunism,” IEEE/ACM Netw. (TON), vol. 18, no. 5, pp. 1464–1477, 2010.
Trans. Netw., vol. 15, no. 5, pp. 1046 –1058, October 2007. [74] A. Garcia-Saavedra, A. Banchs, P. Serrano, and J. Widmer, “Distributed
[47] , “Reducing feedback for opportunistic scheduling in wireless opportunistic scheduling: A control theoretic approach,” in Proc. IEEE
systems,” IEEE Trans. Wireless Commun., vol. 6, no. 12, pp. 4227 INFOCOM, vol. 3, March 2012.
–4232, December 2007. [75] X. Tang, S. Ramprashad, and H. Papadopoulos, “Multi-cell user-
[48] S. Shakkottai, “Effective capacity and QoS for wireless scheduling,” scheduling and random beamforming strategies for downlink wireless
IEEE Trans. Autom. Contr., vol. 53, no. 3, pp. 749 –761, April 2008. communications,” in Proc. IEEE VTC, September 2009, pp. 1 –5.
[49] P. Key, L. Massoulie, A. Bain, and F. Kelly, “Fair internet traffic [76] R. Bendlin, Y.-F. Huang, M. Ivrlac, and J. Nossek, “Fast distributed
integration: network flow models and analysis,” Annals of Telecom- multi-cell scheduling with delayed limited-capacity backhaul links,” in
munications, vol. 59, pp. 1338–1352, 2004. Proc. IEEE ICC, June 2009, pp. 1 –5.
[50] E. Altman, A. Orda, and N. Shimkin, “Bandwidth allocation for guar- [77] W. Choi and J. Andrews, “The capacity gain from base station
anteed versus best effort service categories,” in Proc. IEEE INFOCOM, cooperative scheduling in a MIMO DPC cellular system,” in Proc.
vol. 2, March 1998, pp. 617 –624 vol.2. IEEE International Symposium on Information Theory, July 2006, pp.
[51] T.-J. Lee and G. de Veciana, “Model and performance evaluation for 1224 –1228.
multiservice network link supporting ABR and CBR services,” IEEE [78] , “Base station cooperatively scheduled transmission in a cellular
Commun. Lett., vol. 4, no. 11, pp. 375 –377, November 2000. MIMO TDMA system,” in Proc. IEEE Information Sciences and
[52] S. Borst and P. Whiting, “Dynamic channel-sensitive scheduling algo- Systems, March 2006, pp. 105 –110.
rithms for wireless data throughput optimization,” IEEE Trans. Veh. [79] , “The capacity gain from intercell scheduling in multi-antenna
Technol., vol. 52, no. 3, pp. 569 – 586, May 2003. systems,” IEEE Trans. Wireless Commun., vol. 7, no. 2, pp. 714 –725,
[53] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whiting, and February 2008.
R. Vijayakumar, “Providing quality of service over a shared wireless [80] H. Skjevling, D. Gesbert, and A. Hjørungnes, “Low-complexity dis-
link,” IEEE Commun. Mag., vol. 39, no. 2, pp. 150 –154, February tributed multibase transmission and scheduling,” EURASIP J. Advances
2001. in Signal Processing, vol. 2008, pp. 116:1–116:9, January 2008.
[54] C. Hartmann, R. Vilzmann, A. Schmitt-Nilson, and J. Eberspacher, [81] S. Kiani, D. Gesbert, J. Kirkebo, A. Gjendemsjo, and G. Oien, “A
“Channel aware scheduling for user-individual QoS provisioning in simple greedy scheme for multicell capacity maximization,” in Pro-
wireless systems,” in Proc. IEEE VTC, vol. 2, September 2004, pp. ceedings of International Telecommunications Symposium, September
1009 – 1013 Vol. 2. 2006, pp. 435 –440.
[55] NS2. [Online]. Available: http://www.isi.edu/nsnam/ns/ [82] S. Kiani, G. Oien, and D. Gesbert, “Maximizing multicell capacity
[56] R. Jain, D. Chiu, and W. Hawe, “A quantitative measure of fairness using distributed power allocation and scheduling,” in Proc. IEEE
and discrimination for resource allocation in shared computer systems,” WCNC, March 2007, pp. 1690 –1694.
DEC research report TR-301, 1984. [83] M. Kountouris and D. Gesbert, “Memory-based opportunistic multi-
[57] Z. Zhang, Y. He, and E. Chong, “Opportunistic downlink scheduling user beamforming,” in Proc. IEEE International Symposium on Infor-
for multiuser OFDM systems,” in Proc. IEEE WCNC, vol. 2. IEEE, mation Theory, September 2005, pp. 1426 –1430.
2005, pp. 1206–1212. [84] R. Bartle, The elements of integration and Lebesgue measure. Wiley
[58] X. Liu, E. Chong, and N. Shroff, “A framework for opportunistic Online Library, 1995.
scheduling in wireless networks,” Computer Networks, vol. 41, no. 4, [85] S. Bodas, S. Shakkottai, L. Ying, and R. Srikant, “Scheduling in multi-
pp. 451–474, 2003. channel wireless networks: rate function optimality in the small-buffer
[59] H. Kuhn, “The Hungarian method for the assignment problem,” Naval regime,” in Proc. ACM SIGMETRICS, 2009, pp. 121–132.
research logistics quarterly, vol. 2, no. 1-2, pp. 83–97, 1955. [86] B. Sadeghi, V. Kanodia, A. Sabharwal, and E. Knightly, “Opportunistic
[60] J. Yang, Z. Yifan, W. Ying, and Z. Ping, “Average rate updating media access for multirate ad hoc networks,” in Proc. ACM MobiCom.
mechanism in proportional fair scheduler for HDR,” in Proc. IEEE ACM, 2002, pp. 24–35.
GLOBECOM, 2004, pp. 3464 – 3466 Vol.6. [87] NS3. [Online]. Available: http://www.nsnam.org/
[61] J.-T. Tsai, “State-dependent proportional fair scheduling algorithms for [88] OPNET. [Online]. Available: http://www.opnet.com/
wireless forward link data services,” in Proc. IEEE INFOCOM, April [89] S. Ali, K. Lee, and V. Leung, “Dynamic resource allocation in OFDMA
2008, pp. 2414 –2422. wireless metropolitan area networks [radio resource management and
[62] P. Bender, P. Black, M. Grob, R. Padovani, N. Sindhushyana, and protocol engineering for IEEE 802.16],” IEEE Wireless Commun.,
S. Viterbi, “CDMA/HDR: a bandwidth efficient high speed wireless vol. 14, no. 1, pp. 6–13, 2007.
data service for nomadic users,” IEEE Commun. Mag., vol. 38, no. 7, [90] A. Balasubramanian, R. Mahajan, and A. Venkataramani, “Augmenting
pp. 70 –77, July 2000. mobile 3G using WiFi,” in Proc. ACM MobiSys. ACM, 2010, pp.
[63] A. Jalali, R. Padovani, and R. Pankaj, “Data throughput of CDMA- 209–222.
HDR a high efficiency-high data rate personal communication wireless [91] B. Han, P. Hui, and A. Srinivasan, “Mobile data offloading in
system,” in Proc. IEEE VTC, vol. 3, 2000, pp. 1854 –1858 vol.3. metropolitan area networks,” ACM SIGMOBILE Mobile Computing
[64] R. Almatarneh, M. Ahmed, and O. Dobre, “Performance analysis and Communications Review, vol. 14, pp. 28–30, November 2010.
of proportional fair scheduling in OFDMA wireless systems,” in [92] M. Uysal, Cooperative communications for improved wireless network
Proceedings of Vehicular Technology Conference (VTC). IEEE, 2010, transmission: framework for virtual antenna array applications. In-
pp. 1–5. formation Science Publishing, 2010.
[65] Z. Shen, J. Andrews, and B. Evans, “Adaptive resource allocation in [93] E. Hossain, D. Kim, and V. Bhargava, Cooperative cellular wireless
multiuser OFDM systems with proportional rate constraints,” IEEE networks. Cambridge Univ Pr, 2011.
Trans. Wireless Commun., vol. 4, no. 6, pp. 2726–2737, 2005. [94] J. Laneman, D. Tse, and G. Wornell, “Cooperative diversity in wireless
[66] J.-A. Kwon, B.-G. Kim, and J.-W. Lee, “A unified framework for networks: Efficient protocols and outage behavior,” IEEE Transactions
opportunistic fair scheduling in wireless networks: a dual approach,” on Information Theory, vol. 50, no. 12, pp. 3062–3080, 2004.
Wireless Networks, vol. 16, pp. 1975–1986, October 2010. [95] S. Cui, A. Goldsmith, and A. Bahai, “Energy-efficiency of mimo and
[67] D. Bertsekas, A. Nedi, A. Ozdaglar et al., “Convex analysis and cooperative mimo techniques in sensor networks,” IEEE J. Sel. Areas
optimization,” 2003. Commun., vol. 22, no. 6, pp. 1089–1098, 2004.
[68] D. Bertsekas, “Nonlinear programming,” 1999. [96] S. Jayaweera, “Virtual mimo-based cooperative communication for
[69] C. Kelley, Solving nonlinear equations with Newton’s method. Society energy-constrained wireless sensor networks,” IEEE Trans. Wireless
for Industrial Mathematics, 2003, vol. 1. Commun., vol. 5, no. 5, pp. 984–989, 2006.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
18 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, ACCEPTED FOR PUBLICATION

[97] J. Choi, “Opportunistic beamforming with single beamforming matrix Vincenzo Mancuso obtained his master degree in
for virtual antenna arrays,” in Proc. IEEE International Conference on Electronics from University of Palermo, Italy, in
communications, 2011, pp. 1–5. 2001, and a PhD in Electronics, Computer Science
[98] Q. Deng and A. Klein, “Relay selection in cooperative networks with and Telecommunications from the same University
frequency selective fading,” EURASIP J. Wireless Communications and in 2005. After the PhD, he has collaborated with
Networking, vol. 2011, no. 1, p. 171, 2011. University of Roma “Tor Vergata” and University
[99] K. Choi, E. Hossain, and D. Kim, “Cooperative spectrum sensing of Palermo. He has been visiting scholar at the ECE
under a random geometric primary user network model,” IEEE Trans. Department of Rice University, Houston, Texas, and
Wireless Commun., vol. 10, no. 6, pp. 1932–1944, 2011. postdoc in the MAESTRO team at INRIA Sophia
[100] K. Akkarajitsakul, E. Hossain, and D. Niyato, “Cooperative packet Antipolis, France. Since September 2010, Vincenzo
delivery in hybrid wireless mobile networks: A coalitional game is Staff Researcher at Institute IMDEA Networks,
approach,” IEEE Trans. Mobile Computing, no. 99, pp. 1–1, 2012. working on analytical and experimental projects on wireless networks (802.11
and 802.16/LTE) and energy efficient network protocols. He’s also associate
professor at University Carlos III of Madrid since 2011.

Arash Asadi received a master with highest dis-

tinction in Telematics Engineering from University
Carlos III of Madrid, Spain, in October 2012, and
a master degree in Telecommunication Engineering
from Multimedia University, Malaysia in January
2011. He obtained his bachelor degree in Electronic
Engineering from Islamic Azad University, Iran, in
2006. Since 2011, he has worked as research offi-
cer at Multimedia University, focusing on channel
capacity improvements of wireless networks. Since
July 2011. Arash is with Institute IMDEA Networks,
pursuing his PhD studies while being involved in research project addressing
resource allocation in wireless network and next-generation cooperative
networks.