IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO.

2, APRIL 2011

447

Scheduling Algorithms for Multicarrier Wireless

Data Systems
Matthew Andrews and Lisa Zhang

AbstractWe consider the problem of scheduling multicarrier

wireless data in systems such as IEEE 802.16 (WiMAX). Each
scheduling decision involves assigning carriers to users for each
time slot, subject to the constraint that each carrier is assigned to
at most one user, but multiple carriers can potentially be assigned
to the same user. One important aspect of our problem is that a
scheduler knows the channel rates across all users and all carriers
whenever a scheduling decision is made. This global information
may give a potential for enhancing performance via an optimized
allocation of carriers to users. We analyze this problem in a
situation where nite queues are fed by a data arrival process. The
well-known MaxWeight algorithm for the single-carrier setting
maximizes the product of queue size and service rate. We focus on
how to adapt MaxWeight to the multicarrier setting. If the same
objective is pursued, more service than needed may be assigned to
drain a queue, thereby creating wastage. While a simple variant
in the objective forbids this wastage, it turns an easy-to-compute
old objective into an intractable new objective. We state the
hardness of the new optimization problems and propose several
extremely simple algorithms with provable performance bounds.
We conclude with supporting simulation examples.
Index TermsCommunication systems, communications technology, max weight, multicarrier, scheduling, stability, wireless
communication, wireless networks, wireless systems.

I. INTRODUCTION

HE ADVENT of wireless data systems has led to renewed

interest in scheduling data in multiuser systems. In recent
years, a large body of work has looked at the problem of scheduling over time-varying user-dependent channels in a cellular
wireless system. (See Fig. 1.) This work examines a number of
different models. For example, in the nite-queue model (e.g.,
[29] and [30]), the aim is to keep the system stable, assuming
the queues are fed by an exogenous arrival process. Alternatively, in the innitely backlogged model (e.g., [19], [28], and
[31]), the aim is to maximize the system utility, assuming the
queues are permanently backlogged. Other work examines the
difference between models where the channel rates are governed by some stationary stochastic process and models where a
worst-case adversarial channel process is assumed, e.g., [2] and
[7]. However, most of the previous work looks at a situation
Manuscript received November 12, 2009; revised June 04, 2010; accepted
July 27, 2010; approved by IEEE/ACM TRANSACTIONS ON NETWORKING Editor
T. Bonald. Date of publication September 16, 2010; date of current version April
15, 2011. An earlier version of this paper was published in the Proceedings
of the ACM 13th Annual International Conference on Mobile Computing and
Networking (MobiCom), Montreal, QC, Canada, September 914, 2007.
The authors are with Bell Labs, Murray Hill, NJ 07974 USA (e-mail:
andrews@research.bell-labs.com; ylz@research.bell-labs.com).
Digital Object Identier 10.1109/TNET.2010.2064175

Fig. 1. Cellular wireless system.

with a single wireless carrier in which we can make a scheduling decision on a time slot by time slot basis. Some wireless
systems, however, have multiple carriers in which we can assign different carriers to different users. Examples include multicarrier CDMA systems and also systems such as IEEE 802.16
(WiMAX), EV-DO Revision C, and the Long-Term Evolution
(LTE) of UMTS that use an orthogonal frequency-division multiple access (OFDMA) physical layer in which different tones
can be assigned to different users at each time.
In this paper, we study the problem of scheduling in a timeslotted multicarrier system. A straightforward approach for multiple carriers is to schedule carriers one by one independently by
using an existing scheduling algorithm for each carrier in turn.
Under such a simple adaptation, it is unclear a priori if the performance of a scheduling algorithm can be directly translated
from a single-carrier system to a multicarrier system. For example, all carriers could favor the same user, which could lead
to an excessive amount of service to one user. A main goal of
this paper is to examine how to adapt the popular algorithm
known as MaxWeight to the case of multiple carriers. We present
a number of natural analogs of MaxWeight in the multicarrier
setting and prove their performance bounds against different objective functions. We show the tradeoff between the complexity
of the variants and their performance.
Our approach is based on the natural assumption that a multicarrier scheduler knows the channel rates across all users and all
carriers whenever a scheduling decision is made. This global
information may give a potential for enhancing performance via
an optimized allocation of carriers to users. Another purpose
of this paper is to investigate the benets of jointly allocating
multiple carriers versus the isolated local optimization of each
carrier.
A. Model, Problems, and Results
of

We consider a single base station transmitting data to a set

wireless users on a set of carriers. Due to the wireless

1063-6692/$26.00 2010 IEEE

448

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 2, APRIL 2011

Fig. 2. Schedule during a single time slot t. Shaded squares indicate

x(i; c; t) = 1, and unshaded ones indicate x(i; c; t) = 0.

nature of the channel, the channel rates are user-dependent, carrier-dependent, and time-varying. In particular, we use
to denote the channel rate at which carrier can serve user at
time slot . Knowing the global rate information, a scheduler decides the carrier-to-user assignment during each time slot, subject to the constraint that multiple carriers may be assigned to
the same user, but each carrier however can be assigned to at
most one user. In particular, we use indicator variable
to indicate whether or not carrier is assigned to user at time
. If
, then data of size
can be transmitted
to user at time on carrier (see Fig. 2). Our goal is to choose
in the most advantageous way.
the values
We consider a nite-queue model with an external arrival process. In the single-carrier situation, the well-studied
MaxWeight algorithm always serves the user that maximizes
at each time slot . Here,
denotes the queue
size of user at the beginning of time slot . The MaxWeight algorithm is known to have desirable stability properties, namely
it keeps queues bounded whenever possible. The proof relies
on showing that if the queue sizes are large, then MaxWeight
.
creates a negative drift in the Lyapunov function
The focus of this paper is how to emulate the MaxWeight algorithm in multicarrier systems. Let
be the amount of service user receives at time . We
consider three objective functions when scheduling time slot
(1)
(2)
(3)
Objective (1) is the simplest analog of MaxWeight that
for the single-carrier case.
maximizes
However, optimizing (1) has the potential shortcoming of assigning more service to a user than it can actually use. Note that
this problem exists for the single-carrier setting. However, it is
particularly acute for multicarrier systems where the amount of
service that can be assigned in a single time slot is relatively
large. Objective (2) offers a natural x that forbids excessive
with
in
assignment by replacing
the objective function. Objective (3) explicitly maximizes
and
the negative drift of the Lyapunov function, where
denote the queue size of user at the beginning and at
the end of time slot . Both objectives (2) and (3) are more

sensitive to maintaining small queues than objective (1). We

propose four variations of the MaxWeight algorithm and refer
to them as MaxWeight-Alg1 through MaxWeight-Alg4 (or Alg1
through Alg4 in short). In this paper, we prove a collection of
results regarding these four algorithms in relation to the three
objectives.
1) We describe MaxWeight-Alg1, which assigns each carrier
to the user that maximizes
. This carrier-bycarrier algorithm optimizes objective (1).
2) Somewhat surprisingly, both objectives (2) and (3) are
NP-hard to optimize. Furthermore, they cannot be approxfor some constant .1
imated to within a factor of
3) Since we cannot hope for optimum solutions to objectives
(2) and (3), we focus on developing approximation algorithms. We present MaxWeight-Alg2 and MaxWeight-Alg3,
which are variations of MaxWeight-Alg1. They provide a
-approximation and a -approximation for objectives (2)
and (3), respectively.
In addition, there are scenarios for which MaxWeight-Alg2
fraction of their respective
and Alg 3 achieve at most a
optimal objective values.
4) We present a more complex algorithm MaxWeight-Alg4
that is based on an algorithm for the Generalized Assignment Problem [14]. It improves the approximation ratio for
for any
. Although we
objective (2) to
believe that Alg4 is not simple enough to be practical for
wireless systems, we feel that it does provide theoretical
insight into the multicarrier scheduling problem.
5) We show that the stability properties of the single-carrier
MaxWeight algorithm also apply to the multicarrier algorithms MaxWeight-Alg1 through Alg4.
6) We present simulation results to show that although
MaxWeight-Alg2 and Alg3 may not optimize objectives (2) and (3), they still signicantly outperform
MaxWeight-Alg1 due to the fact that they are trying to
optimize better objectives. The reason for the improved
performance is that MaxWeight-Alg1 often tries to assign
more service to a user than it can actually use. This behavior does not occur for algorithms MaxWeight-Alg2 and
Alg3.
B. Related Work
The MaxWeight algorithm was rst shown to perform well
in wireless networks by Tassiulas and Ephremides [29], [30].
Other papers that study MaxWeight include [4], [3], and [24].
Two algorithms that are similar to MaxWeight are MaxDelay
[4], [3] and Exp [25], [26]. MaxDelay allocates service to user
, where
denotes the head-of-line
delay for user at time . Exp is a more complex algorithm that
provides more control over the relative delays that the users
experience.
These algorithms were designed for the case in which the nite queues are fed by an arrival process. For the case in which
1If opt is the optimal value of a maximization problem, we say an algorithm
is an -approximation algorithm if it always returns a solution whose value is
at least opt. If for every algorithm, there are instances for which the algorithm
cannot guarantee an -approximation, then we say that the problem cannot be
approximated to within a factor .

ANDREWS AND ZHANG: SCHEDULING ALGORITHMS FOR MULTICARRIER WIRELESS DATA SYSTEMS

the queues are innitely backlogged and we wish to maximize a

system utility function, the Proportional Fair algorithm was introduced by [17] and [31] and studied in [1], [19], and [28]. It
was shown in [2] that Proportional Fair does not work so well
when the queues are fed by an arrival process. In particular, it
can cause the queues to be unstable. Algorithms for optimizing
utility functions subject to fairness requirements and constraints
on minimum/maximum throughput have been studied by [5],
[21], and [22]. Algorithms that combine the goals of system
utility maximization and queue stability have recently been presented by [12], [23], and [27]. In [6], it was shown that unless these problems are studied jointly, system oscillations can
occur. We remark that all of this previous work on wireless
scheduling has looked at a single carrier in isolation.
II. CANDIDATE ALGORITHMS
In this section, we dene a number of algorithms that aim to
emulate the MaxWeight algorithm in the multicarrier scenario.
We focus on constructing a schedule for a single time slot. For
convenience, the dependence on is omitted
queue size for user at the beginning of time
which includes the arrival
for time
queue size for user at the end of time
rate for user carrier during time
total service to user during time
queue size for user after carrier is assigned.
Two equations that relate the above quantities are

Recall the objectives (1)(3) dened in Section I. We analyze the

following four algorithms with respect to these three objectives.
The rst three algorithms go through the carriers in order. At
time , carrier serves user dened below.
, where argmax
MaxWeight-Alg1:
.
means is the index that maximizes
.
MaxWeight-Alg2:
MaxWeight-Alg3:
.
MaxWeight-Alg4 begins by approximately solving a relaxation of an integer linear program for objective (2) followed
by rounding the fractional approximate solution. We defer
the detailed description to Section V.
We conclude this section with two simple observations.
The following theorem follows directly from the denition of
MaxWeight-Alg1.
Theorem 1: MaxWeight-Alg1 optimizes objective (1). Our
second result shows that objectives (2) and (3) are related.
Lemma 2: Any -approximation algorithm for objective (2)
provides an -approximation for objective (3).
Proof: Since
, we have

449

In addition,
. Therefore,
objectives (2) and (3) are always within a factor of 2 of each
other.
III. HARDNESS OF OBJECTIVES (2) AND (3)
As we discussed in the Introduction, optimizing objective (1)
is not ideal since it could lead to more service being allocated
to a user than it is able to use, and hence the queue sizes may
become larger than necessary. Hence, it would be preferable to
use objectives (2) and (3). In this section, we show that, unfortunately, we cannot hope for an efcient algorithm that optimizes
objectives (2) and (3) in general.
, there is no
-approximaTheorem 3: For some
.
tion algorithm for objectives (2) and (3) unless
Proof: We use a reduction from the three-bounded threedimensional matching problem. In this problem, we are given
, where
.A
a set
is a subset
such that
three-dimensional matching
no elements in
agree in any coordinate. In a three-bounded
appears at most three times
instance, each element in
of maximum cardinality.
in . The goal is to nd a matching
Kann [18] showed that there exists an such that it is NP-hard
to decide whether the maximum size matching equals or is at
. Specically, given any 3SAT instance, we can
most
construct a matching instance such that if the 3SAT instance is
satisable, then the matching instance has a solution of size ,
and the matching size is at most
otherwise. Without loss
for the constructed matching
of generality, we assume
trivially implies the unsatisability of
instance since
3SAT.
We now convert this matching instance (which is a result
of a reduction from 3SAT) into an instance of our scheduling
problem. We use a reduction similar to that of [11] for a problem
known as the Generalized Assignment Problem. For each hy, we are given a user . For each element
peredge
, we have a carrier . We call these
carriers regif is a
ular carriers. We set the channel rate
component of , and
otherwise. We have another
dummy carriers for which
for
set of
for all users .
all users . Let
Given a scheduling solution, we partition the users into three
, and . Each user in is assigned three regular carsets
riers only. Note that the users in correspond to a three-dimen. Each user in is assigned
sional matching, and hence
one dummy carrier and possibly one regular carrier; each user
in is assigned one or two regular carriers only. To see that
, and form a partition of the users that receive service,
we observe that there is no benet to assigning one dummy carregular carriers to a user since
. There is
rier and
also no benet to assigning two dummy carriers to a user since
there is more benet to reassigning one dummy carrier to a user
regular carriers. Such a user always exists since the
with
is at least
, the number of
number of users in
dummy carriers. Therefore,
and
.
regular carriers not assigned to users in
Consider the
. With respect to objectives (2) and (3), there is more benet to
assigning them to users in than assigning them to users in .
However, we can assign at most two regular carriers to each user

450

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 2, APRIL 2011

in . Hence, at least
regular carriers
are assigned to users in . Therefore, objectives (2) and (3) can
be upper-bounded as follows. In this proof, we use OBJ2 and
OBJ3 to denote objectives (2) and (3). We also use the subscript
to denote the case in which the maximum size matching
to denote the case
has size and the subscript
in which the maximum size matching has size at most

(4)
and

(5)
We now consider two cases. If the size of the maximum threeand
.
dimensional matching is indeed , then
In this case, the upper bounds on objectives (2) and (3) that we
equals (4)
have just derived are actually tight, i.e.,
equals (5). If the maximum three-dimensional
and
, then
. For both
matching has size at most
objectives, the drop in value is at least
, i.e.,
and
.
since the matching instance is threeWe note that
bounded. Therefore, both objectives (2) and (3) are at most
.
This means that the relative difference in the objective values
. By setting
,
between the two cases is at least
we obtain our result.

and
, then there exists an element
such that
.
A special case of a matroid is a partition matroid. We say
that a matroid is a partition matroid if there is a partition of
into components
such that
if and only
, for all .
if
be a function on sets in .
Let
It is a submodular function on if for all
such
and
that
If

It is a nondecreasing submodular function if in addition

and for all
such that

The Greedy algorithm for maximizing a nondecreasing

submodular function over a matroid works as follows.
.
Initially let
Repeat the following procedure for as long as possible.
. Set
Let
.
For partition matroids, the algorithm can be simplied. At
step , instead of considering all elements in , we only
.
need to nd
Fisher et al. [13] proved the following property of the
Greedy algorithm.
Lemma 5: The Greedy Algorithm gives a - approximation
to the problem of maximizing a nondecreasing submodular
function over a matroid.
We now show that MaxWeight-Alg2 is special case of the
Greedy Algorithm for partition matroids. The ground set
. A subset
if and only if
there is at most one element in for each carrier. In other words,
denes a valid schedule. Let
. This
clearly denes a partition matroid. The function
is dened
by

IV. APPROXIMATION RATIOS OF ALG2 AND ALG3

In this section, we show that algorithms MaxWeight-Alg2 and
Alg3 are constant-factor approximations for objectives (2) and
(3). The hardness results of Section III imply that for these objectives, constant-factor approximation algorithms are the best
that we can hope for. Moreover, in Section IV we present simulation results to show that although these algorithms may not
optimize objectives (2) and (3), they still signicantly outperform MaxWeight-Alg1 due to the fact that MaxWeight-Alg1 will
often try to assign more service to a user than it can actually use.
Theorem 4: MaxWeight-Alg2 is a -approximation algorithm
for objective (2). By Lemma 2, this immediately implies that it
is a -approximation algorithm for objective (3).
Proof: We show thatMaxWeight-Alg2 is a special case of
the greedy algorithm for maximizing a nondecreasing submodular function over a matroid. In order to clarify this relationship,
we rst dene the following terms.
Consider a ground set , and let be a set of subsets of .
.
The set is a matroid if
If
and
, then
.

where
, then
any element

. Note that if
. From this, it is easy to see that for
where
forms a valid schedule,
, i.e., the function
is submodular. Moreover, it is clear that function
corresponds directly to objective (2). Hence, when we try to optimize objective (2), we are trying to nd an assignment (which
corresponds to an element of a partition matroid) that maximizes a submodular function. Recall that MaxWeight-Alg2 goes
through each carrier in turn and assigns it to the user that maximizes the increase in the objective. Hence, MaxWeight-Alg2
corresponds to the Greedy algorithm, and so, by the result of
Fisher et al., it is a -approximation algorithm for objective (2).
We now provide an analysis of algorithm MaxWeight-Alg3.
Theorem 6: MaxWeight-Alg3 is a -approximation for
objective (3).

ANDREWS AND ZHANG: SCHEDULING ALGORITHMS FOR MULTICARRIER WIRELESS DATA SYSTEMS

Proof: Recall that

is the queue size for user after algorithm MaxWeight-Alg3 has assigned carrier . Let OPT be
be the analogous queue size
an optimal assignment, and let
after the rst carriers are assigned according to OPT. Consider any carrier . Suppose that OPT assigns carrier to user ,
and MaxWeight-Alg3 assigns it to user . We rst show that the
gain obtained by algorithm MaxWeight-Alg3 due to carrier is
at least one half of the gain obtained by OPT due to carrier that
is never obtained by MaxWeight-Alg3. More precisely, we show

451

by Inequality

since

for

(6)
This immediately implies
Suppose that

. We have

by denition of

since change in queue size bounded by

which completes the proof.

We conclude this section by showing that our analysis
of MaxWeight-Alg2 is essentially tight and our analysis of
MaxWeight-Alg3 cannot be signicantly improved.
, there exists an instance
Theorem 7: For any constant
on which MaxWeight-Alg2 and Alg3 achieve at most a
fraction of the optimal value of objectives (2) and (3).
Proof: The example is as follows. There are two users,
. The channel rates are given by
each with
for
, and
. The optimal
algorithm assigns carrier 1 to user 2 and carrier 2 to user 1.
and
Hence, the optimal values of objectives (2) and (3) are
, respectively. On the other hand, algorithms MaxWeightAlg2 and MaxWeight-Alg3 both assign carrier 1 to user 1 since
.

by assumption that
V. MAXWEIGHT-ALG4: IMPROVED APPROXIMATION
FOR OBJECTIVE (2)
Inequality (6) also holds true for the case that
since
then there is no gain obtained by OPT that is not obtained by
MaxWeight-Alg3. Algebraically, the inequality holds trivially
in this case since the right-hand side is zero. Note that the in.
equality also holds when
Now that we have veried inequality (6), we proceed to prove
the lemma. For clarity of the rest of the proof, we let be the
user that carrier serves under Alg3, and let be the user that
serves under OPT. Note that can be the same as . We know
that
for
, and
for
. We
therefore have

by telescoping on
since

for

In this section, we show that for any

, it is actually
possible to obtain a randomized
-approximation for
.) The algorithm,
objective (2). (Note that
which we call MaxWeight-Alg4, is based on a recent algorithm
for the Generalized Assignment Problem (GAP) due to Fleischer et al. [14]. (In the GAP problem, we are given a set of bins
of different sizes. Each item has a bin-dependent prot and a
bin-dependent size. The goal is to pack the items into bins so
as to maximize the prot in such a way that no bin size is violated.) MaxWeight-Alg4 is somewhat complex, and so we feel
that it is impractical for scheduling wireless systems. However,
we include it here since we feel that it is of theoretical interest
to understand what are the limits regarding the approximability
of objective (2).
be the set of all possible subsets of carriers
Let
, let
that could be assigned to user . For
. For convenience, we also calculate a new set of rates
such that
and
. This can clearly be done in
is used to indicate whether or not
linear time. The variable
subset is assigned to user . We could optimize objective (2)

452

by solving the following integer program:

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 2, APRIL 2011

fraction of the optimal solution to objective (2), our nal ap.

proximation ratio is at least
VI. STABILITY OF FOUR VARIANTS OF MAXWEIGHT

Note that since we showed in Theorem 3 that optimizing objective (2) is NP-hard, we cannot hope to solve the above integer
program exactly. Algorithm MaxWeight-Alg4 nds an approximate solution as follows. It rst nds a solution to the linear
relaxation of the above integer program in which the constraint
is replaced by
. Note that we cannot
directly use a standard linear programming algorithm for this relaxation since there are exponentially many variables. However,
following [14], we can apply standard iterative Lagrangian LP
-approximation to the
algorithms (e.g., [15]) to obtain a
, and
linear relaxation of Alg4-IP in time polynomial in
.
MaxWeight-Alg4 then rounds the solution to this linear relaxation by choosing a single set to assign to user . In particular,
with probability
. We still do not have a valid
we set
assignment since a carrier may be assigned to two different
users. In this case, we pick the user that gives the maximum
.
value of
Theorem 8: The value of the solution obtained by
fraction of the
MaxWeight-Alg4 is at least a
optimal value of objective (2).
to be the
Proof: For each carrier , we set
. We set
user-set pair that has the highest value of
to be the user-set pair that has the next highest value
, etc. The denition of our rounding algorithm
of
as part of
with
means that carrier is assigned to user
.
probability at least
The contribution of carrier in the solution to the relaxation
. By the above argument, the
of Alg4-IP is
expected contribution in the rounded solution is at least

Fleischer et al. [14] show, using the arithmetic/geometric mean

inequality, that expressions of this form are at least

Hence, we have obtained an assignment whose value with refraction of the solution
spect to objective (2) is at least a
to the fractional relaxation. Since this is in turn at least a

In this section, we consider the stability of the four variants

of MaxWeight that we have proposed. Informally, an algorithm
is said to be stable if it keeps the queue sizes bounded. However,
bounded queue is impossible if trafc arrival is unrestricted. We
therefore dene admissible trafc for which some carrier assignment can keep queues bounded. In particular, we use the
-admissible, where
is a window size and
notion of
measures how close the incoming trafc is to full
be the amount of data injected for user in time
load. Let
-admissible if there
slot . For given and , a system is
such that in any window
of
is a schedule
size , we have

Obviously, more injections are admissible to a system with a

larger value or a smaller value. We say that an algorithm
is stable if it keeps the queues bounded as long as the trafc is
-admissible.
The single-carrier MaxWeight algorithm is known to be stable
as long as the channel rates for a user cannot be zero for arbitrarily long periods. (This condition holds for example when
the rates are governed by a stationary stochastic process with
nonzero mean.) The following theorem states that this property
also holds for the four multicarrier MaxWeight algorithms presented in this paper. The proof uses a standard technique (e.g.,
[29] and [30]) of showing that the Lyapunov function
has a negative drift when the queues are large.
and
, algorithms
Theorem 9: If any xed
-admissible
MaxWeight-Alg1 through Alg4 are stable for
trafc as long as for each
the rates
cannot be zero
for arbitrarily long periods.
Proof: We assume that the channel rates are bounded.
be the supremum of these rates. For simplicity, we
Let
also assume that the channel rates are bounded away from
be the inmum of these rates. By looking over
zero. Let
larger time windows, it is straightforward to adapt the argument to the case where channel rates can be zero, but only
for bounded periods of time. Consider the potential function
. We rst show that if the queues are
sufciently large, then this potential function has negative
drift for MaxWeight-Alg1. Recall that
indicates whether or not carrier serves user at time . If
, then

ANDREWS AND ZHANG: SCHEDULING ALGORITHMS FOR MULTICARRIER WIRELESS DATA SYSTEMS

and otherwise

453

This implies

In both cases this implies

Hence, when the queues are sufciently large, i.e., larger than
, the potential function decreases.
This implies stability for MaxWeight-Alg1.
For algorithms MaxWeight-Alg2 and Alg4, stability follows
from a similar argument to the above and the fact that if
, then each of these algorithms assigns each carrier to
. Hence, in all cases
user

Since the arrival process is

-admissible, the total arrivals
, and .
per user is upper-bounded by a function of
In particular

Similarly, the second term can be upper-bounded by a function

of
, and . Let denote an upper bound of the rst
two terms. To bound the third term, we note that for

where is a function of
, and .
, then algorithm MaxWeight-Alg3
Similarly, if
for some
assigns carrier to user
. Therefore, we once again have

Hence
VII. SIMULATIONS

Again, the second term of the above expression can be upperand . Let
be this conbounded by a function of
stant. To bound the third term, let
for given time slot and carrier . The denition of max weight
and
for all
. Therefore
implies

Given and , let

. In other words,
. Therefore, given and , we have

We now present simulation work. Our focus is on the rst

three MaxWeight algorithms, as their simplicity is more likely
to lead to practical consequences, whereas we feel the signicance of Alg4 is more theoretical. The question we address is
whether we are better off using an optimal algorithm, namely
MaxWeight-Alg1, for a not-so-good objective function or
an approximate algorithm, for example MaxWeight-Alg2 or
MaxWeight-Alg3, for a more sensible objective function. Our
nding chooses the latter.
The algorithms are implemented in simple homegrown programs written in Python. We use a eld trace that represents
measured channel conditions from a third-generation wireless
system as well as synthetic traces in which the channel rates
uctuate around a mean value according to 3 km/h Rayleigh
fading. We assume a constant rate arrival model. The number of
carriers is 40, and the number of users varies between 30 and 40.
In all cases, MaxWeight-Alg2 and Alg3 have extremely similar
performance, and so we combine them onto a single plot. Both
of these algorithms signicantly outperform MaxWeight-Alg1.
This is due to the fact that MaxWeight-Alg1 is wasteful and often

454

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 2, APRIL 2011

Fig. 3. Simulated trace: total queue size. (Left) MaxWeight-Alg1. Mean:

119440; 95%: 141900; Median: 118904. (Right) Alg2 and Alg3. Mean: 4275;
95%: 8888; Median: 3736.

Fig. 5. Field trace: total queue size over all users per time step.
(Left) MaxWeight-Alg1. Mean: 140746; 95%: 167828; Median: 131448.
(Right) Alg2 and Alg3. Mean: 226; 95%: 260; Median: 22.

Fig. 4. Simulated trace: total queue size over all users per time step.
(Left) MaxWeight-Alg1. Mean: 30247; 95%: 41328; Median: 29488. (Right)
Alg2 and Alg3. Mean: 31; 95%: 176; Median: 0.

Fig. 6. Simulated trace: delay averaged over packets that complete transmission at each time step. (Left) MaxWeight-Alg1. (Right) Alg2 and Alg3.

assigns more service to a user than it can actually use. See Figs. 3
and 4 for total queue size plots under the simulated traces, and
Fig. 5 for plots under the eld trace. The gure captions offer
summary statistics for the mean, 95th percentile, and median
queue sizes.

VIII. CONCLUSION
In this paper, we studied a variety of scheduling algorithms
for time-slotted multicarrier wireless data systems. We presented a set of algorithms that aim to emulate the MaxWeight
algorithm for the single-carrier case.
A number of immediate open problems remain. First, we
would like to know if it is possible to improve on the -approximation for objective (3). In particular, we would like to know
if algorithm MaxWeight-Alg3 has a better approximation ratio
than since, in the worst example that we can construct, the
performance of MaxWeight-Alg3 only differs from the optimum
by a factor of . We would also like to know if there is a simple

algorithm that improves on the -approximation for objective

(2). We feel that MaxWeight-Alg4 is probably too complex to be
implemented in practice.
There are also other longer term open problems such as deriving delay bounds for scheduling problems in the multiuser
multicarrier setting. The relationship between small queues and
small delays is however intuitive, illustrated by Fig. 6, whose
queue-size counterpart is plotted in Fig. 3.
Since the preliminary version of this work appeared in ACM
MobiCom in 2007, more work has been done on multicarrier
scheduling, e.g., [10] presents an iterative longest-queue-rst
algorithm that aims to minimize buffer overow probabilities.
A number of new research problems stemming from multicarrier wireless scheduling has been dened. One line of research
focuses on accommodating a contiguity constraint, which is imposed by the standards for the uplink of LTE. This constraint
requires that if multiple carriers are assigned to the same user
during a time slot, then these carriers have to be consecutive.
Approximation algorithms for this problem are proposed in [8]
and [20].

ANDREWS AND ZHANG: SCHEDULING ALGORITHMS FOR MULTICARRIER WIRELESS DATA SYSTEMS

REFERENCES
[1] R. Agrawal and V. Subramanian, Optimality of certain channel aware
scheduling policies, in Proc. 40th Annu. Allerton Conf. Commun.,
Control, Comput., Monticello, IL, Oct. 2002, pp. 15321541.
[2] M. Andrews, Instability of the proportional fair scheduling algorithm for HDR, IEEE Trans. Wireless Commun., vol. 3, no. 5, pp.
14221426, Sep. 2004.
[3] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, R. Vijayakumar,
and P. Whiting, CDMA data QoS scheduling on the forward link
with variable channel conditions, Bell Labs, Tech. Memorandum, Apr.
2000.
[4] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, R. Vijayakumar,
and P. Whiting, Providing quality of service over a shared wireless
link, IEEE Commun. Mag., vol. 39, no. 2, pp. 150154, Feb. 2001.
[5] M. Andrews, L. Qian, and A. Stolyar, Optimal utility based multi-user
throughput allocation subject to throughput constraints, in Proc. IEEE
INFOCOM, 2005, vol. 4, pp. 24152424.
[6] M. Andrews and A. Slivkins, Oscillations with TCP-like ow control
in networks of queues, in Proc. IEEE INFOCOM, 2006, pp. 112.
[7] M. Andrews and L. Zhang, Scheduling over non-stationary wireless
channels with nite rate sets, IEEE/ACM Trans. Netw., vol. 14, no. 5,
pp. 10671077, 2006.
[8] M. Andrews and L. Zhang, Multiserver scheduling with contiguity
constraints, in Proc. IEEE INFOCOM, Rio, Brazil, Apr. 2009, pp.
12781286.
[9] P. Bender, P. Black, M. Grob, R. Padovani, N. Sindhushayana, and A.
Viterbi, CDMA/HDR: A bandwidth efcient high speed data service
for nomadic users, IEEE Commun. Mag., vol. 38, no. 7, pp. 7077,
Jul. 2000.
[10] S. Bodas, S. Shakkottai, L. Ying, and R. Srikant, Scheduling in multichannel wireless networks: Rate function optimality in the small-buffer
regime, in Proc. SIGMETRICS, Seattle, WA, 2009, pp. 121132.
[11] C. Chekuri and S. Khanna, A PTAS for the multiple knapsack
problem, in Proc. 11th Annu. ACM-SIAM Symp. Discrete Algor., San
Francisco, CA, Jan. 2000, pp. 213222.
[12] A. Eryilmaz and R. Srikant, Fair resource allocation in wireless networks using queue-length based scheduling and congestion control, in
Proc. IEEE INFOCOM, Miami, FL, Mar. 2005, vol. 3, pp. 17941803.
[13] M. Fisher, G. Nemhauser, and L. Wolsey, An analysis of approximations for maximizing submodular set functions-II, Math. Program.
Studies, vol. 8, pp. 7387, 1978.
[14] L. Fleischer, M. Goemans, V. Mirrokni, and M. Sviridenko, Tight approximation algorithms for maximum general assignment problems,
in Proc. 17th Annu. ACM-SIAM Symp. Discrete Algor., Miami, FL,
2006, pp. 611620.
[15] N. Garg and J. Knemann, Faster and simpler algorithms for multicommodity ow and other fractional packing problems, in Proc.
39th Annu. Symp. Found. Comput. Sci., Palo Alto, CA, Nov. 1998, pp.
300309.
[16] D. Hochbaum, Various notions of approximations; good, better, best,
and more, in Approximation Algorithms for NP-Hard Problems.
Florence, KY: Course Technology, 1995, pp. 346398.
[17] A. Jalali, R. Padovani, and R. Pankaj, Data throughput of
CDMA-HDR a high efciency-high data rate personal communication wireless system, in Proc. IEEE VTC2000-Spring, Tokyo,
Japan, May 2000, vol. 3, pp. 18541858.
[18] V. Kann, Maximum bounded 3-dimensional matching is max SNPcomplete, Inf. Process. Lett., vol. 37, no. 1, pp. 2735, 1991.
[19] H. Kushner and P. Whiting, Asymptotic properties of proportionalfair sharing algorithms, in Proc. 40th Annu. Allerton Conf. Commun.,
Control, Comput., 2002, pp. 10511059.
[20] S. Lee, I. Pefkianakis, A. Meyerson, S. Xu, and S. Lu, Proportional fair
frequency-domain packet scheduling for 3GPP LTE uplink, in Proc.
IEEE INFOCOM, Rio, Brazil, Apr. 2009, pp. 26112615.

455

[21] X. Liu, E. Chong, and N. B. Shroff, Opportunistic transmission scheduling with resource-sharing constraints in wireless networks, IEEE J.
Sel. Areas Commun., vol. 19, no. 10, pp. 20532064, Oct. 2001.
[22] X. Liu, E. Chong, and N. B. Shroff, A framework for opportunistic
scheduling in wireless networks, Comput. Netw., vol. 41, no. 4, pp.
451474, 2003.
[23] M. Neely, E. Modiano, and C. Li, Fairness and optimal stochastic control for heterogeneous networks, in Proc. IEEE INFOCOM, Miami,
FL, Mar. 2005, vol. 3, pp. 17231734.
[24] M. Neely, E. Modiano, and C. Rohrs, Power and server allocation
in a multi-beam satellite with time varying channels, in Proc. IEEE
INFOCOM, New York, Jun. 2002, vol. 3, pp. 14511460.
[25] S. Shakkottai and A. Stolyar, Scheduling algorithms for a mixture of
real-time and non-real-time data in HDR, in Proc. 17th ITC, Salvador
da Bahia, Brazil, 2001, pp. 793804.
[26] S. Shakkottai and A. Stolyar, Scheduling for multiple ows sharing
a time-varying channel: The exponential rule, Anal. Methods Appl.
Probab., vol. 207, pp. 185202, 2002.
[27] A. Stolyar, Maximizing queueing network utility subject to stability:
Greedy primal-dual algorithm, Queueing Syst., vol. 50, no. 4, pp.
401457, 2005.
[28] A. Stolyar, On the asymptotic optimality of the gradient scheduling
algorithm for multiuser throughput allocation, Oper. Res., vol. 53, pp.
1225, 2005.
[29] L. Tassiulas and A. Ephremides, Stability properties of constrained
queueing systems and scheduling policies for maximum throughput in
multihop radio networks, IEEE Trans. Autom. Control, vol. 37, no. 12,
pp. 19361948, Dec. 1992.
[30] L. Tassiulas and A. Ephremides, Dynamic server allocation to parallel
queues with randomly varying connectivity, IEEE Trans. Inf. Theory,
vol. 39, no. 2, pp. 466478, Mar. 1993.
[31] D. Tse, Multiuser diversity in wireless networks, Dept. EECS, UC
Berkeley, Berkeley, CA, 2001 [Online]. Available: http://www.eecs.
berkeley.edu/~dtse/stanford416.ps

Matthew Andrews received the B.A. degree (rst class honors) in mathematics
from Oxford University, Oxford, U.K., in 1993, and the Ph.D. degree in theoretical computer science from the Massachusetts Institute of Technology, Cambridge, in 1997.
He is a Distinguished Member of Technical Staff with the Mathematics of
Networks and Systems Department, Bell Labs, Murray Hill, NJ. He holds a
range of patents in the area of telecommunications. His research interests include wireless resource allocation, packet scheduling, and approximation algorithms. His recent work includes algorithms for joint scheduling and congestion
control in ad hoc networks and complexity theoretic results on the hardness of
network design.

Lisa Zhang received the B.A. degree (summa cum laude) in mathematics from
Wellesley College, Wellesley, MA, in 1993, and the Ph.D. degree in theory of
computing from the Massachusetts Institute of Technology, Cambridge, in 1997.
She is a Member of Technical Staff with the Algorithms Research Group,
Bell Labs, Murray Hill, NJ. She holds a range of patents in the area of telecommunications. Her research area is algorithm design and analysis. Her research
broadly concerns algorithmic and complexity issues of networking, with a focus
on design and optimization, routing and scheduling protocols, and stability and
quality-of-service analyses.
Dr. Zhang twice won the Bell Labs Presidents Gold Award and the Lucent
Chairmans Award.