IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP.

1146-1159, AUGUST 2009 1

Delay Analysis for Maximal Scheduling with Flow
Control in Wireless Networks with Bursty Trafc
Michael J. Neely
AbstractWe consider the delay properties of one-hop net-
works with general interference constraints and multiple trafc
streams with time-correlated arrivals. We rst treat the case when
arrivals are modulated by independent nite state Markov chains.
We show that the well known maximal scheduling algorithm
achieves average delay that grows at most logarithmically in
the largest number of interferers at any link. Further, in the
important special case when each Markov process has at most two
states (such as bursty ON/OFF sources), we prove that average
delay is independent of the number of nodes and links in the
network, and hence is order-optimal. We provide tight delay
bounds in terms of the individual auto-correlation parameters
of the trafc sources. These are perhaps the rst order-optimal
delay results for controlled queueing networks that explicitly
account for such statistical information. Our analysis treats cases
both with and without ow control.
Index TermsQueueing analysis, Markov chains, Flow Control
I. INTRODUCTION
This paper derives average delay bounds for one-hop
wireless networks that use maximal scheduling subject to
a general set of interference constraints. It is known that
maximal scheduling algorithms are simple to implement and
can support throughput within a constant factor of optimality.
Our analysis shows that this type of scheduling also yields
tight delay guarantees. In particular, when arrival processes
are modulated by independent Markov processes, we show that
average delay grows at most logarithmically in the number of
nodes in the network. We then obtain an improved delay bound
in the important special case when the individual Markov
chains have at most two states (such as bursty ON/OFF
sources). Average delay in this case is shown to be independent
of the network size, and hence is order-optimal.
Specially, we consider a network with N nodes and L
links. Let N and L denote the node and link sets:
N

=
{1, 2, . . . , N} , L

=
{1, 2, . . . , L}
Each link l L represents a directed communication channel
from one node to another, and we dene tran(l) and rec(l) to
be the corresponding transmitter and receiver nodes for link l
(where tran(l) N and rec(l) N). The network operates
in slotted time with unit timeslots t {0, 1, 2, . . .}. Every
timeslot a decision is made about which links to activate for
transmission. If a link l is activated during a particular slot,
This work was presented in part at the IEEE INFOCOM conference,
Phoenix, AZ, April 2008.
Michael J. Neely is with the Electrical Engineering department at the
University of Southern California, Los Angles, CA (web: [Link]
[Link]/mjneely).
This material is supported in part by one or more of the following: the
DARPA IT-MANET program grant W911NF-07-0028, the NSF grant OCE
0520324, the NSF Career grant CCF-0747525.
it sends exactly one packet from tran(l) to rec(l). However,
due to scheduling and/or interference constraints, not all links
can be simultaneously active during the same slot. These
constraints are dened according to the general interference
model of [2][3]: Each link l L is allowed to be active if
and only if no other links within an interference set S
l
are
simultaneously active. For convenience, it is useful to dene
the set S
l
to additionally include link l itself. That is, each
set S
l
consists of link l together with all possible interferers
of link l. Note that these interference sets have the following
pairwise symmetry property: For any two links , l L, we
have that S
l
if and only if l S

.
The link sets S
l
can be chosen to impose a variety of
constraint models. For example, setting S
l
to include all
links adjacent to either the transmitter or receiver of link
l imposes matching constraints. Matching constraint models
arise naturally in scheduling problems for packet switches.
They are also important for wireless networks where individual
nodes can transmit or receive over at most one adjacent link,
and cannot simultaneously transmit and receive (called the
node exclusive interference model in [4]). More general sets
S
l
can be used to model topology-dependent interference
constraints for wireless networks, such as the constraint that
no additional transmitting nodes can be activated within a
specied distance of a node that is actively receiving.
Every timeslot, new data randomly arrives to the network.
Let A
l
(t) represent the (integer) number of packets that arrive
during slot t that are intended for transmission over link
l. Packets are stored in separate queues according to their
corresponding link. This is a one-hop network, so that packets
exit the network once they are transmitted over their intended
link. A network controller observes the current queue backlog
and makes link activation decisions every slot subject to the
transmission constraints.
It is well known that generalized max-weight scheduling can
be used to achieve maximum throughput in such networks [5]
[6]. However, this type of scheduling is difcult to implement
in wireless networks with general interference constraints. In
this paper we consider a simpler class of maximal scheduling
algorithms. Maximal scheduling is of recent interest due to
its low complexity and ease of distributed implementation.
For N N packet switches, maximal scheduling is known to
support throughput that is within a factor of 2 of optimality,
and to also have nice delay properties for i.i.d. inputs [7]
[8]. Related constant factor throughput results have also been
shown for wireless networks, including factor of 2 results for
networks under matching constraints [4] and constant factor
results for more general interference models [2] [3]. However,
the work on wireless scheduling in [4][2][3] considers only
throughput results and does not provide a delay analysis.
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 2
Further, while the work in [8] considers delay analysis for i.i.d.
arrivals and a NN packet switch, no existing work provides
explicitly computable and order-optimal delay bounds for
time-correlated arrivals.
Our work addresses the issues of general interference con-
straints and time-correlated bursty trafc simultaneously.
We treat the general interference model of [2][3], but use
the concept of queue grouping to derive the order-optimal
delay results. Queue grouping techniques have been used
in [9][10][7][3] to reduce scheduling complexity in switches
and wireless networks. They are used in [11] to provide
order-optimal delay for opportunistic scheduling in a single-
server downlink. Near order-optimal delay is established for
N N packet switches in [12]. The analysis in this paper
particularly treats delay in wireless networks with general
constraint sets and time-correlated arrivals. Previous work in
[13][14][6] treats time-correlated arrivals, but does not obtain
order optimal delay. Asymptotic delay optimality is studied in
[15] for a heavy trafc regime where input rates are scaled so
that they are very close to the capacity region boundary. Here,
we focus on the case when input rates are a xed fraction away
from the boundary. We obtain tight delay bounds in terms
of the individual auto-correlation parameters of the trafc
sources. These are perhaps the rst delay bounds for controlled
queueing networks that explicitly incorporate such statistical
information. This allows delay to be understood in terms of
general models for network trafc. Our analysis includes the
important special case of Markov modulated ON/OFF trafc
sources, allowing for explicit delay bounds in terms of the
parameters of each source.
We rst treat the case of general Markovian arrivals and
prove a structural result about average delay, showing that
average delay grows at most logarithmically in the worst case
number of interferers of a given link, and hence is at most
O(log(N)). We then provide an explicit and tighter delay
analysis for the special case when all Markov chains have
at most two states. In this case, we prove average delay is
independent of N. Our analysis rst assumes arrival rates are
inside the stability region associated with maximal scheduling,
and is next generalized to treat ow control when trafc rates
are either inside or outside of this region. The time-correlated
scenarios treated here are quite challenging to analyze, and we
introduce a simple technique of delayed Lyapunov analysis
to ensure the arrival processes couple sufciently fast to a
stationary distribution. We note that this delayed Lyapunov
analysis is different from the related T-slot drift technique
developed in [13] for queue stability with Markovian channels
(also used in [14][6] for both stability and delay analysis), as
the T-slot drift approach cannot achieve order-optimal delay
results.
In the next section we present the network model. In
Section III we provide the drift analysis, and in Section IV we
present the logarithmic delay result for general time-correlated
arrivals. The order-optimal delay analysis for 2-state chains is
provided in Section V. Flow control mechanisms for systems
with arbitrary trafc rates are considered in Section VI.
II. NETWORK MODEL
Recall that L denotes the set of network links, and that
transmission over each link is constrained by the general
interference sets dened in the previous section.
A. Trafc Model
Suppose the arrival process A
l
(t) is modulated by a dis-
crete time, stationary, ergodic Markov chain Z
l
(t) for each
link l L. Specically, Z
l
(t) has nite state space Z
l
=
{1, 2, . . . , M
l
}. For each link l L and state m Z
l
, arrivals
A
l
(t) are conditionally independent and identically distributed
according to mass function p
(m)
l
(a), where:
p
(m)
l
(a) = Pr[A
l
(t) = a | Z
l
(t) = m] for a {0, 1, 2, . . .}
Dene the conditional arrival rates
(m)
l
as follows:

(m)
l
= E{A
l
(t) | Z
l
(t) = m}
We assume conditional second moments of arrivals are nite,
so that E
_
(A
l
(t))
2
| Z
l
(t) = m
_
< for all l L and all
m Z
l
. Let
(m)
l
represent the steady state probability that
Z
l
(t) = m. Dene
l
as the average arrival rate to link l:

l
=

mZ
l

(m)
l

(m)
l
(1)
We assume that all Markov chains are in their steady state
distribution at time 0, so that each A
l
(t) process is stationary
and for all slots t 0 and all links l L we have:
E{A
l
(t)} =
l
The Markov chains Z
l
(t) themselves may be correlated
over different links l L, although we mainly focus on the
case when chains are independent. More detailed statistical
information, such as the auto-correlation for individual inputs
(and the spatial correlation between multiple inputs if they are
not independent), is also important for delay analysis and shall
be dened when needed. Note that this trafc model is quite
general and includes the following important special cases:
Case 1: Z
l
(t) has only one state and so arrivals A
l
(t)
are i.i.d. over slots with some given distribution.
Case 2: Z
l
(t) is a 2-state ON/OFF process where A
l
(t) =
1 whenever Z
l
(t) = ON and A
l
(t) = 0 whenever
Z
l
(t) = OFF.
Let
2
l

=
E
_
A
l
(t)
2
_

2
l
represent the steady state arrival
variance for link l. While our results hold for general trafc
with nite variance, it is important to note that the order-
optimal delay analysis we achieve requires the following
assumption:
1

tot

lL

2
l
= O(1) (2)
where
tot

lL

l
. This is a mild assumption that typically
holds for any inputs with a nite variance. For example,
if arrivals are Poisson then we have
2
l
=
l
, and so
1
tot

lL

2
l
= 1. Similarly, it is easy to show that if there is
a nite constant A
max
such that A
l
(t) A
max
for all l and
all t, then
1
tot

lL

2
l
A
max
.
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 3
B. Queueing
Dene Q
l
(t) as the number of queued packets waiting for
transmission over link l during slot t. Let Q(t) = (Q
l
(t))
lL
be the vector of queue backlogs. Dene
l
(t) {0, 1} as the
transmission rate offered to the link during slot t (in units
of packets/slot). That is,
l
(t) = 1 if link l is scheduled for
transmission on slot t, and
l
(t) = 0 otherwise. We assume
the scheduler only schedules a link l that does not violate
the interference constraints and that has a packet ready for
transmission (so that Q
l
(t) > 0). Let (t) = (
l
(t))
lL
represent the transmission rate vector for slot t. Dene X(t) as
the set of feasible transmission vectors for slot t, representing
all (t) rate vectors that conform to the constraints dened
by the interference sets S
l
and the additional constraint that

l
(t) = 1 only if Q
l
(t) > 0 (for each l L). The queueing
dynamics thus proceed as follows:
Q
l
(t + 1) = Q
l
(t)
l
(t) +A
l
(t) (3)
The goal is to observe the queue backlogs every slot and
make scheduling decisions (t) X(t) so as to support all
incoming trafc with average delay as small as possible.
C. Maximal Scheduling
Dene the network capacity region as the closure of
the set of all arrival rate vectors (
l
)
lL
that can be stably
supported, considering all possible scheduling algorithms that
conform to the above constraints (see [6] for a discussion of
capacity regions and stability). It is well known that scheduling
according to a generalized max-weight rule every timeslot
ensures stability and maximum throughput whenever arrival
rates are interior to the capacity region [5] [6].
1
However, the
max-weight rule involves an integer optimization that may be
difcult to implement, and has delay properties that are dif-
cult to analyze. Here, we assume scheduling is done according
to a simpler maximal scheduling algorithm. Specically, given
a queue backlog vector Q(t), a transmission vector (t) is
maximal if it satises the interference constraints and is such
that for all links l L, if Q
l
(t) > 0 then

(t) = 1 for
at least one link S
l
. In words, this means that if link l
has a packet, then either link l is selected for transmission,
or some other link within the interference set S
l
is selected.
There is much recent interest in maximal scheduling because
of its implementation simplicity (described briey below) and
its ability to support input rates within a constant factor of
the capacity region for wireless networks [4] [2] [3] and for
N N packet switches [7].
One way to achieve a maximal scheduling is as follows:
First select any non-empty link l L and label it active.
Then select any other non-empty link that does not conict
with the active link l (i.e., that is not within S
l
). Label this
second link active. Continue in the same way, selecting
new non-empty links that do not conict with any previously
selected links, until no more links can be added. It is not
difcult to see that this nal set of links labeled active has
1
Specically, the generalized max-weight rule in this case schedules to
maximize
P
lL
Q
l
(t)
l
(t) subject to (t) X(t).
the desired maximal property. Maximal link selections are not
unique, and can alternatively be found in a distributed manner,
where multiple nodes attempt to activate their non-conicting,
non-empty links simultaneously, and contentions are resolved
locally. This distributed implementation also requires multiple
iterations before the set of selected links becomes maximal.
All maximal link selections have the following important
mathematical property.
Lemma 1: Under any maximal link scheduling for (t), for
all links l L we have:
Q
l
(t)

S
l

(t) Q
l
(t) (4)
Proof: Consider any link l L. If Q
l
(t) = 0, then (4)
reduces to 0 0 and is trivially true. Else, if Q
l
(t) > 0 then

(t) = 1 for at least one link within S

l
(by denition of
a maximal link selection), and so

S
l

(t) 1, which
proves (4).
In this paper, we assume transmission decisions are made
every slot according to any maximal scheduling. For conve-
nience, we further assume that the maximal scheduling has a
well dened probabilistic structure given the queue backlog
vector, so that the entire queueing system can be viewed as
an ergodic Markov chain with a countably innite state space.
The inequality (4) is the only additional property of maximal
scheduling required in our analysis.
D. The Reduced Throughput Region
Dene

as the set of all rate vectors (

l
)
lL
that satisfy
the following:

S
l

1 for all l L
The set

is overly restrictive, as it is possible to have more

than one simultaneously active link within a given set S
l
(provided that link l is idle). However, the set

is typically
within a constant factor of the capacity region . For example,
in networks with matching constraints only, it is not difcult
to show that
1
2

, so that the throughput region

is
within a factor of 2 of optimality. Further, in networks with
general inteference sets S
l
where each set S
l
can support at
most K simultaneously active links (called the interference
degree of the network), it can be shown that
1
K

[2].
It is easy to see in this case that the integer K is strictly less
than the largest cardinality of any interference set. The work
in [2] also constructs a network for which the set
1
K
shares
a common boundary point with the set of all rates supportable
through maximal scheduling.
Throughout this paper, we assume input rates (
l
)
lL
are
interior to the set

. Specically, we assume there exists a

value

such that 0 <

< 1, where:

S
l

for all l L (5)

The value

represents the relative network loading, as it

can be viewed as a loading factor relative to the reduced
throughput region

.
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 4
E. An Example for N N Packet Switches
For an illuminating example of the reduced throughput re-
gion

and its comparison to the capacity region , consider

a NN packet switch operating under the crossbar constraint
with input rate matrix (
ij
) for input ports i {1, . . . , N} and
output ports j {1, . . . , N} (for a more detailed discussion of
packet switches and the crossbar constraint, see, for example,
[16] [17] [12]). There are N
2
links, each labeled (i, j) ac-
cording to its corresponding input/output pair. Link activation
sets correspond to matchings on the resulting bi-partite graph.
Specically, if link (i, j) is activated for transmission, then no
other link (i, b) or (a, j) can be activated (for b = j, a = i).
Therefore, the interference set S
ij
for each link (i, j) is:
S
ij
= {(a, j) | a {1, . . . , N}} {(i, b) | b {1, . . . , N}}
It is well known that the capacity region for this N N
switch is given by the set of all (
ij
) rate matrices that satsify
the following 2N inequality constraints:

N
b=1

ib
1 for all i {1, . . . , N}

N
a=1

aj
1 for all j {1, . . . , N}
On the other hand, the reduced throughput region

is given
by the set of all non-negative rate matrices (
ij
) that satisfy
the following N
2
inequality constraints:
N

a=1

aj
+

b{1,...,N},b=j

ib
1 for all links (i, j)
That is, (
ij
)

if and only if all rates are non-negative,

and for any crosspoint (i, j), the sum of the rates in row i and
column j (including crosspoint (i, j) only once) is less than
or equal to 1. It is not difcult to see that if a matrix in has
all entries halved, then it will be in

, and hence:
1
2

However, the reduced throughput region

is strictly larger
than
1
2
. For example, consider the following rate matrices for
a 2 2 switch and a 3 3 switch, respectively:
(
ij
) =
_
.7 .1
.1 .1
_
, (
ij
) =
_
_
.1 .2 0
.3 .2 0
.3 0 .2
_
_
The rst matrix is outside the set
1
2
because the rst column
sums to 0.8 (larger than 0.5). The second matrix is also outside
of the set
1
2
(for the corresponding 3 3 capacity region )
because the rst column sums to 0.7. However, both matrices
are inside their respective reduced throughput regions

, and
both have relative loading

= 0.9.
III. DRIFT ANALYSIS
Recall that Q(t) = (Q
l
(t)). Our technique relies on the
concept of queue grouping. For each link l L, dene:

Q
S
l
(t)

S
l
Q

(t) (6)
Thus,

Q
S
l
(t) is the sum of all queue backlogs of links within
the interference set S
l
of link l. Dene the Lyapunov function:
L(Q(t))

=
1
2

lL
Q
l
(t)

Q
S
l
(t) (7)
The queue-grouped structure of this Lyapunov function is
similar to the functions used in [3][9] to prove queue stability
when input rates are a xed fraction away from the capacity
boundary. An alternate proof of rate-stability is given in [2].
However, the prior work in this area does not directly consider
delay performance. Below we provide a more detailed drift
analysis that yields explicit and tight delay bounds.
For each link l L, dene the group departures
S
l
(t) and
group arrivals

A
S
l
(t) as follows:

S
l
(t)

S
l

(t) ,

A
S
l
(t)

S
l
A

(t)
Thus:

Q
S
l
(t + 1) =

Q
S
l
(t)
S
l
(t) +

A
S
l
(t) (8)
Dene the 1-step unconditional Lyapunov drift as follows:
(t)

=
E{L(Q(t + 1)) L(Q(t))} (9)
where the expectation is over the randomness of Q(t) and the
randomness of the system dynamics given the value of Q(t).
Lemma 2: (Drift Under Maximal Scheduling) If maximal
scheduling is implemented every timeslot (using any maximal
scheduling algorithm), the resulting unconditional Lyapunov
drift (t) satises the following for all slots t:
(t) E{B(t)}

lL
E
_
Q
l
(t)(1

A
S
l
(t))
_
(10)
where
B(t)

=
1
2

lL
_
A
l
(t)

A
S
l
(t) 2

A
S
l
(t)
l
(t) +
l
(t)
_
(11)
Proof: See Appendix A.
A. Lyapunov Drift Theorem
The drift expression (10) can be used to prove stability
and delay properties of maximal matching via the following
theorem:
Theorem 1: (Lyapunov drift [6]) Let Q(t) be a vector
process of queue backlogs that evolve according to some
probability law, and let L(Q(t)) be a non-negative function
of Q(t). If there exist processes f(t) and g(t) such that the
following is satised for all time t:
(t) E{g(t)} E{f(t)}
then:
limsup
t
1
t
t1

=0
E{f()} limsup
t
1
t
t1

=0
E{g(t)}
The proof of Theorem 1 involves summing the telescoping
series (see [6]).
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 5
B. Analysis for i.i.d. Arrivals
Dene A(t)

=
(A
l
(t))
lL
as the vector of new arrivals.
Consider rst the case when all arrival vectors A(t) are i.i.d.
over timeslots, with rate vector = (
l
)
lL
(the arrivals over
different links in the same slot may be correlated). For each
link l, dene

S
l
as the sum of arrival rates over all input
streams corresponding to links within the set S
l
. That is:

S
l

S
l

Note by the loading assumption (5) that

S
l

, where

is a value such that 0 <

< 1. By independence of arrivals

every slot, we have for all t:
E
_
Q
l
(t)(1

A
S
l
(t))
_
= E{Q
l
(t)} (1

S
l
)
E{Q
l
(t)} (1

)
Using this inequality directly in the Lyapunov drift expression
(10) yields (for all slots t):
(t) E{B(t)} (1

lL
E{Q
l
(t)}
Plugging the above drift inequality into the Lyapunov
drift theorem (Theorem 1) (using g(t)

=
B(t) and f(t)

=
(1

lL
Q
l
(t)) yields:
limsup
t
1
t
t1

=0

lL
E{Q
l
()}
B
1

(12)
where:
B

=
limsup
t
1
t
t1

=0
E{B()}
Note from (11) that B < , and hence the queueing
network is strongly stable with nite time average queue
backlogs. Because it evolves according to an ergodic Markov
chain with a countably innite state space, it can be shown
that limiting time averages exist and are equal to the steady
state averages. Thus, the left hand side of (12) represents the
time average total queue backlog in the system (summed over
all queues). Let Q
l
be the average queue backlog in link l (for
each l L). We thus have:
Theorem 2: (Time-Independent Arrivals) If the arrival vec-
tor process A(t) is i.i.d. over slots with a relative network
loading

< 1 (dened in (5)), then:

(a) The average total network congestion satises:

lL
Q
l

lL
_
E
_
A
l
(t)

A
S
l
(t)
_
2
l

S
l
+
l
_
2(1

)
(b) If arrival streams A
l
(t) are i.i.d. over slots and are
also independent of each other, then W, the expected delay
averaged over all packets in the network, satises:
W
1 +
1
tot

lL
_

2
l

l

S
l
_
2(1

)
(13)
where
tot

lL

l
, and where
2
l

=
E
_
(A
l
(t))
2
_

2
l
and
represents the variance of A
l
(t).
The average congestion bound in part (a) of the above
theorem is found by computing B using (11) and noting that
the system is stable, has a steady state, and that the time
average limit of E{
l
(t)} is equal to
l
. Part (b) is proven
by the fact that total average congestion is equal to
tot
W
(by Littles Theorem). This also uses inter-link independence
for the identity E
_
A
l
(t)

A
S
l
(t)
_
=
2
l
+
l

S
l
. Note that the
numerator in (13) is O(1) if (2) holds.
C. Delay under Poisson and Bernoulli Inputs
Note that if all arrival processes A
l
(t) are independent and
Poisson with rate
l
, we have that
2
l
=
l
. The average delay
bound (13) in this case reduces to:
W
Poisson

1
1
2tot

lL

S
l
(1

)
(14)
This demonstrates that average delay is O(1), that is, it is in-
dependent of the network size N. Hence, maximal scheduling
achieves order optimal delay with respect to N, provided that
the arrival rates are interior to the reduced throughput region

, as described by the constraints (5). This is in contrast to

the O(N) average delay bounds derived for the throughput-
optimal max-weight scheduling for N N packet switches in
[17] and for wireless networks in [14] [18].
2
The expression
in the right hand side of (14) also provides an upper bound on
delay in the case of independent Bernoulli arrivals, because
2
l
for a Bernoulli variable is less than that of a Poisson variable.
IV. LOGARITHMIC DELAY FOR TIME-CORRELATED
ARRIVALS
Consider the case of nite-state ergodic Markov modulated
arrivals, as described in Section II-A. Assume that all trafc
rates satisfy the loading constraints (5) with relative network
loading

. Let H(t) represent the past history of all actual

arrivals (of all processes) up to but not including time t. For a
given link l L, suppose there exists a non-negative function

l
(T) (for T {0, 1, 2, . . .}) such that, regardless of past
history H(t), we have:
E{A
l
(t) | H(t T)}
l
+
l
(T) (15)
and such that:
lim
T

l
(T) = 0
That is,
l
(T) characterizes the time required for the process
A
l
(t) to converge to its stationary mean, regardless of the
initial condition. It can be shown that all nite state ergodic
Markov processes converge exponentially fast to their steady
state (see, for example, [19]). Hence for each l L we can
dene
l
(T) as follows:

l
(T) =
l

T+1
l
(16)
for some constant
l
and some decay factor
l
such that 0 <

l
< 1. The
l
and
l
constants can in principle be determined
2
We emphasize that max-weight scheduling is a special case of maximal
scheduling, and so our delay bounds also apply to max-weight in the case
when arrival rates are inside the reduced throughput region

.
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 6
as parameters from the Markov chain Z
l
(t). Here, we prove
a structural result concerning logarithmic delay in terms of
these parameters. Dene (T) as follows:
(T)

=
max
lL
_

S
l
+

S
l

(T)
_
(17)
Note that

S
l

< 1 for all l L, and so there exist

integers T such that (T) < 1.
Theorem 3: (General Time-Correlated Arrivals) If arrival
processes have rates (
l
)
lL
in the interior of

, then for
any integer T 0 such that (T) < 1, we have:
(a) The average total network congestion satises:

lL
Q
l

B +

F
T
1 (T)
where:

B

=
1
2

lL
_

l
+E
_
A
l
(t)

A
S
l
(t)
__
(18)

F
T

lL
T

k=1
E
_

A
S
l
(k)A
l
(0)
_
(19)
(b) If arrival processes A
l
(t) for different links l are
additionally independent of each other, then the constants

B
and

F
T
satisfy:

B =
1
2

lL
_

l
+
l

S
l
+
2
l
_
(20)

F
T
=

lL
T

k=1
[

S
l

l
+
l
(k)] (21)
where
2
l

=
E
_
(A
l
(t))
2
_

2
l
is the variance of A
l
(t), and

l
(k) is the auto-correlation in A
l
(t) and is dened:

l
(k)

=
E{A
l
(t +k)A
l
(t)}
2
l
Proof: The proof is given in Appendix B.
A. Discussion of the Delay Result
By Littles Theorem, if the conditions of Theorem 3 are
satised, then average network delay W satises:
W
1
tot
(

B +

F
T
)
1 (T)
The parameter T only affects the delay bound and does not
affect the maximal scheduling algorithm. Thus, the bound can
be optimized over all integers T such that (T) < 1. Here
we show how the resulting bound grows as a function of the
network size. First note that in the case when arrival processes
are independent of each other, the constant

B in (20) satises:

B
tot
_
1 +
1
2
tot

lL

2
l
_
where
tot

lL

l
. This is because

S
l
1 for all l L.
Likewise, the constant

F
T
in (21) satises:

F
T

tot
T +

lL
T

k=1

l
(k)
Therefore, when arrival processes A
l
(t) are independent of
each other, average network delay satises:
W
T + 1 +
1
2tot

lL

2
l
+
1
tot

lL

T
k=1

l
(k)
1 (T)
The values of
1
2tot

lL

2
l
and
1
tot

lL

l
(k) are typ-
ically independent of N (recall (2)), and so the numerator is
roughly linear in the T value. Because we have nite state
ergodic Markov chains, from (16) we see the function (T)
has the form:
(T)

+|S|
T+1
where |S| is the cardinality of the largest interference set S
l
,
and and are the largest values of
l
and
l
, respectively,
over all links l L. In this case, we have 1 (T) (1

)/2
whenever:
|S|
T+1
(1

)/2
which holds when T is chosen as the smallest integer that
satises:
log (2 |S| /(1

))
log(1/)
1 T
log (2 |S| /(1

))
log(1/)
Thus, the above delay bound grows at most logarithmically
in |S|. A more explicit and order-optimal delay analysis is
provided in the next section, where the special case of 2-state
Markov chains is considered and average delay is shown to
be independent of the network size.
V. TIME-CORRELATED ARRIVALS WITH TWO STATES
Consider the case of Markov modulated arrivals, as de-
scribed in Section II-A, where all Markov chains Z
l
(t) have
at most two states (labeled 1 and 2). Note that this model
includes the important special case of ON/OFF inputs, where
A
l
(t) has a single packet arrival when in the ON state and has
no arrivals when in the OFF state.
3
Let

L be the set of all links
l L that have exactly two states with different conditional
rates
(1)
l
and
(2)
l
. The transition probabilities are given by

l
and
l
for each two-state chain Z
l
(t), as shown in Fig. 1.
Assume that 0 <
l
< 1 and 0 <
l
< 1 for all l

L, and
dene
(1)
l
and
(2)
l
as the steady state probabilities for each
two-state chain Z
l
(t):

(1)
l
=

l

l
+
l
,
(2)
l
=

l

l
+
l
3
The conditional rates for such an ON/OFF example are given by
(1)
l
= 1,

(2)
l
= 0, where states 1 and 2 are associated with the ON and OFF states,
respectively, as shown in Fig. 1.
1
(ON)
2
(OFF)

l
1-
l
1-
l
Fig. 1. The 2-state Markov chain Z
l
(t) for link l.
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 7
Arrival processes A
l
(t) for the remaining links l /

L are
either i.i.d. over slots (effectively one-state chains), or have
two states but with
(1)
l
=
(2)
l
(in the latter case, the
different states may correspond to different conditional second
moments).
The Markov chains Z
l
(t) are possibly correlated over dif-
ferent links l L, although we focus primarily on the case
when all chains are independent. The chains are assumed to be
stationary, so that for each link l L we have E{A
l
(t)} =
l
for all time t. The time average rates
l
for all 2-state arrival
processes are given by:

l
=
(1)
l

(1)
l
+
(2)
l

(2)
l
The trafc rates (
l
)
lL
are assumed to satisfy the loading
constraints (5) with relative network loading

< 1. In the
previous section we showed that the system is strongly stable
and ergodic with nite steady state queue backlogs. Here we
provide a tighter delay analysis using a combination of Markov
chain theory and Lyapunov drift theory.
A. 2-State Drift Analysis
Recall the denition

A
S
l
(t)

S
l
A

(t). Thus, from the

drift inequality (10) we have that the unconditional Lyapunov
drift (t) satises the following every slot t:
(t) E{B(t)}

lL
E{Q
l
(t)}
+

lL

S
l
E{Q
l
(t)A

(t)} (22)
Now assume that the system is in steady state at time t, and
is also in steady state at time t 1. Fix a link l L and an
arrival process A

(t). We shall derive a relationship between

E{Q
l
(t)A

(t)} and E{Q

l
(t)}. To this end, rst note that for
any link l L we have:
E{Q
l
(t)A

(t)} = E{Q
l
(t)}

if /

L (23)
That is, the expectation has a simple product form in the
special case when /

L, because A

(t) is either i.i.d. over

slots or has the same conditional rate
(1)

=
(2)

.
For the rest of this subsection we consider the opposite and
more challenging case when

L. In this case, we have:
E{Q
l
(t)} =
2

m=1

(m)

E{Q
l
(t) | Z

(t) = m} (24)
However, we also have:
E{Q
l
(t)A

(t)}
=
2

m=1

(m)

E{Q
l
(t)A

(t) | Z

(t) = m}
=
2

m=1

(m)

(m)

E{Q
l
(t) | Z

(t) = m} (25)
This nal equality holds because the expectation of Q
l
(t)
is conditionally independent of A

(t) given Z

(t). Be-
cause the system is in steady state, the quantities E{Q
l
(t)},
E{Q
l
(t)A

(t)}, and E{Q

l
(t) | Z

(t) = m} do not depend

on t, and we dene:
E{Q
l
}

=
E{Q
l
(t)}
E{Q
l
A

}

=
E{Q
l
(t)A

(t)}
x
(m)
l,

=

(m)

E{Q
l
(t) | Z

(t) = m}
The equalities (24) and (25) can be re-written:
E{Q
l
} = x
(1)
l,
+x
(2)
l,
(26)
E{Q
l
A

} =
(1)

x
(1)
l,
+
(2)

x
(2)
l,
(27)
Equations (26) and (27) are two linear equations that express
a relationship between 4 unknowns (where the unknowns are
E{Q
l
} , E{Q
l
A

}, x
(1)
l,
and x
(2)
l,
). To express a direct linear
relationship between E{Q
l
} and E{Q
l
A

}, we require an
additional equation. To this end, note that:
Q
l
(t) = Q
l
(t 1)
l
(t 1) +A
l
(t 1)
Therefore:
E{Q
l
(t)A

(t)} = E{Q
l
(t 1)A

(t)} D
l,
+C
l,
(28)
where C
l,
and D
l,
are dened as:
C
l,

=
E{A
l
(t 1)A

(t)} (29)
D
l,

=
E{
l
(t 1)A

(t)} (30)
Now:
E{Q
l
(t 1)A

(t)}
=
2

m=1

(m)

E{Q
l
(t 1)A

(t) | Z

(t 1) = m}
=
2

m=1

(m)

h
(m)

E{Q
l
(t 1) | Z

(t 1) = m} (31)
where h
(m)

is dened:
h
(m)

=
E{A

(t) | Z

(t 1) = m} (32)
The last equality follows again because Q
l
(t1) is condition-
ally independent of A

(t) given Z

(t 1). However, because

the system is in steady state at time t and also at time t 1,
it follows that:

(m)

E{Q
l
(t 1) | Z

(t 1) = m} = x
(m)
l,
(33)
Therefore, using (33) and (31), equation (28) becomes the
following:
E{Q
l
A

} = C
l,
D
l,
+h
(1)

x
(1)
l,
+h
(2)

x
(2)
l,
(34)
The constants h
(m)

, dened in (32), can be computed directly

from the transition probabilities for chain Z

(t):
h
(1)

= (1

)
(1)

(2)

(35)
h
(2)

(1)

+ (1

)
(2)

(36)
The linear equations (26), (27), and (34) involve three
equations and four unknowns, and can be shown to be linearly
independent whenever
(1)

=
(2)

(which holds for all

IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 8


L). This directly leads to the following lemma that relates
E{Q
l
} and E{Q
l
A

}.
Lemma 3: For all links l L, we have:
E{Q
l
A

} = E{Q
l
}

+
C
l,
D
l,

if

L (37)
Proof: The result follows by eliminating the x
(1)
l,
and x
(2)
l,
variables from (26), (27), and (34), and the computation is
given in Appendix C.
Note from the denitions (29) and (30) that C
l,
0,
D
l,
0 for all l L and all

L. Dene C
l,
= 0
for /

L. Using (37) and (23) we nd that for all link pairs
l, L, we have:
E{Q
l
A

} E{Q
l
}

+
C
l,

where the inequality comes because we have neglected the

D
l,
constant. Using this expression directly in the Lyapunov
drift inequality (22) yields the following drift expression that
holds at any time t at which the system is in steady state:
(t) E{B(t)}

lL
E{Q
l
(t)} (1

)
+

lL

S
l

L
C
l,

(38)
where we have used the fact that

S
l

.
B. The Delay Bound for 2-State Markov Modulated Arrivals
Theorem 4: If the input rates (
l
)
lL
satisfy the loading
constraints (5) for a given relative network loading

< 1,
then:
(a) The system is stable with steady state average congestion
that satises:

lL
Q
l

B +

C
1

where:

B

=
1
2

lL
_
E
_
A
l
(t)

A
S
l
(t)
_
+
l
_

C

=

lL

S
l

L
E{A
l
(t 1)A

(t)}

(b) If all Markov chains Z

l
(t) are additionally independent
of each other, then:

B =
1
2

lL
_

S
l
+
2
l
+
l
_
E{A
l
(t 1)A

(t)} =
l

if l =
Hence, average delay W satises:
W
1 +
1
tot

lL
[
2
l
+
l

S
l
]
2(1

)
+
1
tot
_

l
[1]
(
l
+
l
)
+

S
l

(+)
_
1

where
tot

lL

l
, and
l
[1]

=
E{A
l
(t 1)A
l
(t)} (
l
)
2
is the 1-slot auto-correlation for process A
l
(t), given by:

l
[1] =

l

l
(
(1)
l

(2)
l
)
2
(1
l

l
)
(
l
+
l
)
2
Proof: The result follows directly from (38) via the
Lyapunov drift theorem (Theorem 1). Details omitted for
brevity.
Using the fact that

S
l

< 1, the average delay bound

W in part (b) of the above theorem can be simplied as
follows:
W
1 +

+
1
tot

lL

2
l
2(1

)
+max

L
_
1

_
_

+
1
tot

l
[1]
1

_
Note that
l
[1]
l

(max)
, where
(max)
is the largest
conditional rate over all links and states. Thus, the numer-
ator in the nal term in the above delay bound satises:
1
tot

l
[1]
max
. Therefore, the above bound is O(1)
(independent of the network size N). In the special case of
ON/OFF sources, where a single packet arrives from stream l
when Z
l
(t) = ON and no packet arrives when Z
l
(t) = OFF,
we have
(1)
l
= 1 and
(2)
l
= 0, and
l
=
(1)
l
. Further:

2
l

l
,
l
[1]
l
. Thus, average delay in this ON/OFF
example satises:
W
ON/OFF

1 +

2
+ max

L
[(

+ 1)/(

)]
(1

)
Note that 1/
l
is the average burst size (i.e., the average
time spent in the ON state), and so the numerator roughly
grows linearly in the largest average burst size over any input.
VI. DELAY ANALYSIS WITH FLOW CONTROL
The previous sections assume that the steady state input
rate vector (
l
)
lL
satises the loading contstraints (5), and
hence is interior to the reduced throughput region

. Here we
consider arbitrary input rate vectors (possibly inside or outside
of

), and develop a ow control mechanism that works

together with maximal link scheduling. We restrict attention
to arrival processes A
l
(t) that are Markov modulated and have
at most two Markov states, as in the previous section.
We use the ow control framework of [6]. Specically,
every slot the ow controller at link l observes the current
number of new packets A
l
(t), and decides how many of these
new arrivals to place into the network. Let R
l
(t) be this ow
control decision, where R
l
(t) is an integer that satises:
0 R
l
(t) A
l
(t) for all l L and all t
Any packets that are not immediately admitted by the ow
controller are dropped. The queueing dynamics are thus given
by (compare with (3)):
Q
l
(t + 1) = Q
l
(t)
l
(t) +R
l
(t) (39)
Let (r
l
)
lL
denote the vector of time average admission
rates into the network links (for l L), which corresponds
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 9
to some particular ow control algorithm. We dene network
fairness in terms of concave utility functions g
l
(r). Each utility
function g
l
(r) is continuous, concave, non-negative, and non-
decreasing. The goal is to design a ow control algorithm
that ensures order optimal delay while maximizing the sum of
utilities over all network links, where the maximum is dened
with respect to the reduced throughput region

. Specically,
dene g

as the maximum sum utility associated with the

following optimization problem:
Maximize:

lL
g
l
(r
l
) (40)
Subject to: (r
l
)

(41)
0 r
l

l
for all l L (42)
We desire our ow control algorithm to yield a sum utility
that is close to (or larger than) the value of g

. We note
that a similar utility-based ow control problem with maximal
scheduling is considered in [4], and a token based technique
for max-min fairness is considered in [2]. Our algorithm is
quite different from [4] [2], and our analysis is unique in that
it demonstrates order-optimal delay for bursty trafc arrivals.
For technical reasons, we dene I

as the nite set of real

numbers between 0 and 1, inclusive, that are uniformly spaced
apart by some (arbitrarily small) quantization > 0. For
example, I

might represent numbers as viewed by a computer

that rounds to a nite decimal place. This quantization is
convenient for limit theorems, as we will show the resulting
system has only a nite number of possible states. Dene g

as the maximum value of the problem (40)-(42) subject to the

additional constraint:
r
l
I

for all l L (43)

It is clear that g

is very close to g

when is small. We shall

modify our goal to achieve a target utility of g

, rather than
g

.
A. The Flow Control Algorithm with Maximal Scheduling
Note that because the utility functions are non-decreasing,
the optimization problem (40)-(43) is equivalent to:
Maximize:

lL
g
l
(
l
)
Subject to: (r
l
)

r
l

l
for all l L
0
l
r
l
,
l
I

for all l L
where we have introduced an auxiliary variable
l
for each
link l. Using our ow control framework of [6] [20], for each
link l we dene an auxiliary process
l
(t) and a ow state
queue Y
l
(t). The ow state queue Y
l
(t) for each link l is
implemented purely in software at link l, is initialized so that
Y
l
(0) = 0, and has update equation:
Y
l
(t + 1) = max[Y
l
(t) R
l
(t), 0] +
l
(t) (44)
where R
l
(t) is the admission decision made by link l on slot
t and
l
(t) is an auxiliary variable chosen by link l on slot t
according to the following algorithm. The algorithm is dened
in terms of a control parameter V > 0 that affects a tradeoff
between utility and delay.
Flow Control with Maximal Scheduling (FLOW-MAXIMAL):
Every slot t, the ow controller at each link l L observes
the queue backlog

Q
S
l
(t) (dened in (6)), the ow state queue
Y
l
(t), and the new arrivals A
l
(t), and performs the following:
(Admission Control) Choose R
l
(t) as follows:
R
l
(t) =
_
A
l
(t) if Y
l
(t)

Q
S
l
(t)
0 otherwise
(Auxiliary Variables) Choose
l
(t) as the solution to:
Maximize: V g
l
(
l
(t)) Y
l
(t)
l
(t) (45)
Subject to: 0
l
(t) 1 ,
l
(t) I

(46)
Note that this is a simple maximization of a concave
function of a single variable over an interval. The
additional constraint
l
(t) I

can be enforced by rst

calculating the non-quantized optimal
l
(t) value, and
then choosing the quantized value either to the right or
left according to which one maximizes (45).
(Virtual Queue Update) Update Y
l
(t) according to (44),
using the above chosen values of R
l
(t) and
l
(t).
The network then schedules links as before, using any maximal
scheduling algorithm. Specically, every slot t the following
action is taken:
(Maximal Scheduling) Choose (
l
(t))
lL
by performing
maximal link scheduling, so that queue backlog is up-
dated according to (39).
B. Algorithm Performance
Suppose arrivals A
l
(t) are Markov modulated (possibly
correlated over the different links l), and satisfy the two-state
assumptions of Section V with steady state trafc rate vector
(
l
)
lL
. For simplicity, we additionally assume the maximum
number of packets that can arrive to any link during a slot is
bounded by a constant A
max
, so that:
A
l
(t) A
max
for all t
Consider any utility functions g
l
(r) that are concave, non-
decreasing, and non-negative. We assume that g
l
(0) = 0
for all l L (so that zero throughput yields zero utility).
Further dene
l
as the maximum right-derivative of the utility
function g
l
(r).
4
We assume that
l
< for all l L, and
dene as the largest
l
value:

=
max
lL

l
. For example,
consider the following two types of utility functions dened
in terms of non-negative constants
l
and
l
:
Linear: g
l
(r) =
l
r
Logarithmic: g
l
(r) =
l
log(1 +
l
r)
In the linear case above, we have
l
=
l
. In the logarithmic
case, we have
l
=
l

l
.
Our rst result shows that the above algorithm yields
bounded queue sizes for all time.
Theorem 5: (Worst Case Queue Bounds) Suppose all
queues are initially empty and the above FLOW-MAXIMAL
algorithm is implemented. Then for arbitrary arrival processes
4
All concave functions of one variable have well dened right-derivatives.
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 10
A
l
(t) that satisfy the maximum arrival bound A
l
(t) A
max
for all t, we have the following:
0 Y
l
(t) V
l
+ 1 for all t (47)
0 Q
l
(t) V
l
+ 1 +A
max
for all t (48)
Proof: First note that because g
l
() is concave, has
maximum derivative
l
, and satises g
l
(0) = 0, we have:
g
l
()
l
for all 0
Now suppose that Y
l
(t) > V
l
for some slot t. The
l
(t)
variable is chosen to maximize (45). However, for any possible
value 0, we have:
V g
l
() Y
l
(t) V
l
Y
l
(t) V
l
V
l
= 0
where the second inequality is an equality only if = 0.
It follows that the controller chooses
l
(t) = 0 on any slot
in which Y
l
(t) > V
l
. It follows from (44) that Y
l
(t) cannot
increase on the next slot, and so it can only increase when it is
currently less than or equal to V
l
. Because
l
(t) 1 for all t,
the maximum amount of increase is 1, and so Y
l
(t) V
l
+1
for all t, proving (47).
Similarly, note that if Q
l
(t) > V
l
+ 1, then

Q
S
l
(t) >
V
l
+ 1 Y
l
(t), and so the admission control algorithm sets
R
l
(t) = 0. It follows that Q
l
(t) can only increase when it is
less than or equal to V
l
+1. Because A
max
is the maximum
amount it can increase, we have that (48) holds for all time.
Note from (44) and the constraint
l
(t) I

that Y
l
(t)
can only take values that are integers plus an element in I

.
Because Y
l
(t) is bounded, it can only take a nite number of
values. Thus, there are only a nite number of system states
[Z(t); Q(t); Y (t)] (recall that Q
l
(t) is a bounded integer and
Z(t) has a nite state space). It follows that time average
expectations are well dened for any xed initial condition.
Let (r

l
)
lL
represent an optimal solution to (40)-(43) with
optimal sum utility g

, so that g

=

lL
g
l
(r

l
).
Theorem 6: (Performance of Flow Control with Maximal
Scheduling) Suppose the above FLOW-MAXIMAL algorithm
is implemented. Suppose arrivals are modulated by ergodic
Markov processes with at most two-states as described in
Section V, with steady state rates (
l
)
lL
(possibly inside or
outside of

). Then queue backlog and sum utility satisfy:

lL
Q
l
r
tot
_
1 +
A
max
2
+V
_
+
1
2

lL
E
_
A
l
(t)

A
S
l
(t)
_
(49)

lL
g
l
(r) g

F/V (50)
where Q
l
is the time average expected backlog in link l, and
r
l
is the time average expected admitted rate to link l:
Q
l

=
lim
t
1
t
t1

=0
E{Q()} , r
l

=
lim
t
1
t
t1

=0
E{R
l
()}
and r
tot
=

lL
r
l
. Finally, the constant F is given by:
F

=
r
tot
_
1 +
A
max
2
_
+
1
2

lL
E
_
A
l
(t)

A
S
l
(t)
_
+

lL

S
l

L
(r

)E{A
l
(t 1)A

(t)}

L
(r

l
/
l
)E{A
l
(t 1)A
l
(t)}

l
+
l
(51)
where r

l
/
l
1 due to the constraints of (40)-(42). Recall
that the set

L contains only links that have exactly two states
and satisfy
(1)
l
=
(2)
l
. If the set

L is empty (such as when
arrivals are i.i.d. over slots), then the last two summation terms
in the expression for F are equal to zero.
We emphasize that the E
_
A
l
(t)

A
S
l
(t)
_
value in Theorem
6 is taken with respect to the steady state distribution for
arrivals. Further, because the total input rate into the system is
r
tot
, we have by Littles Theorem that average delay satises:
W
_
1 +
A
max
2
+V +
1
2r
tot

lL
E
_
A
l
(t)

A
S
l
(t)
_
_
It follows that we again have an explicit tradeoff between
utility and average delay, as determined by the V parameter.
Further, if exogenous inputs are independent of each other and

tot
/r
tot
= O(1), we have:
1
2r
tot

lL
E
_
A
l
(t)

A
S
l
(t)
_
= O(1)
In this case, average delay is O(1) (independent of the
network size). Theorem 6 is proven in Appendix D using a
novel argument that combines Lyapunov optimization with the
steady state analysis for bursty trafc as in Section V.
C. Discussion of Periodic Maximal Scheduling
Because any maximal scheduling algorithm can be used, one
can consider algorithms where the initial collection of links
that are activated as part of the maximal set are determined
periodically, so that each link is chosen for this initial activa-
tion every M slots (where M is a function of the network size).
Links in this initial set that contain packets are scheduled for
transmission, and additional links are then chosen to ensure the
maximal property is satised in the network on every slot. For
example, in NN bi-partite graphs with matching constraints,
one can consider choosing links according to a round-robin
schedule that selects one link every N slots. This ensures a
service opportunity every N slots, and hence, because queue
backlog Q
l
(t) is bounded by V
l
+1+A
max
for all time, we
also have a worst-case delay bound of N[V
l
+ 1 + A
max
]
slots. This worst-case delay bound holds in addition to all of
our other analytical guarantees, including our average delay
bound that is independent of N when
tot
/r
tot
= O(1).
VII. CONCLUSION
We have developed order-optimal delay results for one-hop
networks with general interference set constraints and bursty
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 11
(time-correlated) trafc. Our results hold for cases when trafc
is within the reduced throughput region

, a region that
is typically within a constant factor of the network capacity
region. Futher, we have developed a ow control technique
that works together with maximal scheduling and yields an
explicit utility-delay tradeoff.
APPENDIX A PROOF OF LEMMA 2
To compute (t), note that using (8) and (3) yields:
Q
l
(t + 1)

Q
S
l
(t + 1) = Q
l
(t)

Q
S
l
(t)
+(A
l
(t)
l
(t))(

A
S
l
(t)
S
l
(t))
Q
l
(t)(
S
l
(t)

A
S
l
(t))

Q
S
l
(t)(
l
(t) A
l
(t))
Thus, the 1-step unconditional Lyapunov drift is given by:
(t) = E{B(t)}

1
2

lL
E
_
Q
l
(t)(
S
l
(t)

A
S
l
(t))
_

1
2

lL
E
_

Q
S
l
(t)(
l
(t) A
l
(t))
_
(52)
where
B(t)

=
1
2

lL
_
(A
l
(t)
l
(t))(

A
S
l
(t)
S
l
(t))
_
(53)
We now use the following important structural property of
the interference sets.
Lemma 4: (Sum Switching) For any function f(l, )
(where l L, L), we have:

lL

S
l
f(l, ) =

lS
f(l, )
The above lemma follows directly from the pairwise symmetry
property of the interference sets: For any two links l,
L, we have that S
l
if and only if l S

, and hence
{(l, ) | l L, S
l
} = {(l, ) | L, l S

}. Using this
lemma we can re-write the nal term in (52):

lL

Q
S
l
(t)(
l
(t) A
l
(t))
=

lL

S
l
Q

(t)(
l
(t) A
l
(t)) (54)
=

lS
Q

(t)(
l
(t) A
l
(t)) (55)
=

L
Q

(t)(
S
(t)

A
S
(t)) (56)
=

lL
Q
l
(t)(
S
l
(t)

A
S
l
(t)) (57)
where (54) follows by the denition of

Q
S
l
(t) given in (6),
(55) follows by the Sum Switching Lemma (Lemma 4), and
(57) follows by re-labeling the indices. Plugging the equality
(57) directly into the drift expression (52) yields:
(t) = E{B(t)}

lL
E
_
Q
l
(t)(
S
l
(t)

A
S
l
(t))
_
(58)
Further, we note that the expression for B(t) in (53) is
equivalent to that given in (11). This can be seen by using a
sum-switching argument similar to (54)-(57) on the summation

lL

S
l
(t)A
l
(t), and by noting that
l
(t)
S
l
(t) =
l
(t).
The latter equality holds because if
l
(t) = 0 we have 0 = 0
which is trivially true, while if
l
(t) = 1 then no other links
within S
l
can be active, and so
S
l
(t) = 1.
We now use the fact that maximal scheduling is per-
formed every timeslot. Specically, we recall that any maximal
scheduling algorithm satises (4) every timeslot. Using the
denition of
S
l
(t), we note that (4) is equivalent to:
Q
l
(t)
S
l
(t) Q
l
(t) for all l L and all t
Plugging the above inequality directly into (58) yields the
expression (10) for (t) under maximal scheduling.
APPENDIX B PROOF OF THEOREM 3
To prove Theorem 3, we introduce an articial delay in the
nal term of the drift expression (10) to decouple correlations
between queue state and arrivals. This is similar to the T-
slot technique of [13][14][6], although, unlike [13][14][6], it
allows a tight logarithmic delay result. To begin, x an integer
T 0, and note that for t {0, 1, 2, . . . , } we have:
Q
l
(t) Q
l
(t T) +
T1

v=0
A
l
(t T +v)
where we dene Q
l
(t) = 0, A
l
(t) = 0 for t < 0. Using the
above inequality in (10) yields
(t) E{B(t) +F
T
(t)}

lL
E{Q
l
(t)}
+

lL
E
_
Q
l
(t T)

A
S
l
(t)
_
(59)
where F
T
(t) is dened:
F
T
(t)

lL
E
_

A
S
l
(t)
T1

v=0
A
l
(t T +v)
_
We now use the (T) function to modify the nal term on
the right hand side of (59):
E
_
Q
l
(t T)

A
S
l
(t)
_
= E
_
Q
l
(t T)E
_

A
S
l
(t) | Q(t T)
__
E
_
Q
l
(t T)
_

S
l
+

S
l

(T)
__
(60)
E{Q
l
(t T)} (T) (61)
where (60) follows from (15) and (61) follows from (17).
Using (61) in (59), it follows that unconditional Lyapunov
drift satises:
(t) E{B(t) +F
T
(t)}

lL
E{Q
l
(t)}
+ (T)

lL
E{Q
l
(t T)} (62)
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 12
Fixing the integer T and using the Lyapunov drift theorem
(Theorem 1) in the drift expression (62) yields:
limsup
t
1
t
t1

=0

lL
E{Q
l
()}
_
B +F
T
_
+ (T) limsup
t
1
t
t1

=0

lL
E{Q
l
( T)} (63)
where
B

=
limsup
t
1
t
t1

=0
E{B()} , F
T

=
limsup
t
1
t
t1

=0
E{F
T
()}
However, it is not difcult to see that:
limsup
t
1
t
t1

=0

lL
E{Q
l
( T)}
= limsup
t
1
t
t1

=0

lL
E{Q
l
()} (64)
Indeed, the equality (64) follows by noting the time-delayed
version of the limit on the left hand side does not affect the
overall time average, as any contribution to the sum over the
extra T slots is nite and becomes negligible as t .
Therefore, because it is assumed that (T) < 1, the inequality
(63) simplies to:
limsup
t
1
t
t1

=0

lL
E{Q
l
()}
B +F
T
1 (T)
(65)
It is easy to show that:
B
1
2

lL
_

l
+E
_
A
l
(t)

A
S
l
(t)
__
F
T
=

lL
E
_

A
S
l
(t)
T1

v=0
A
l
(t T +v)
_
It follows by stationarity that B

B and F
T
=

F
T
, where

B and

F
T
are dened in (18) and (19). Finally, we note that
because the queueing dynamics are described by a countable
state space, irreducible Markov chain, the limsup in the left
hand side of (65) can be replaced by a regular limit, which
proves part (a) of Theorem 3. Part (b) of Theorem 3 follows
by computing

B and

F
T
for the case when arrival processes
A
l
(t) are independent of each other.
APPENDIX C PROOF OF LEMMA 3
Choose any two links l, such that l L and

L.
The linear equations (27) and (34) can be re-written in matrix
form:
E{Q
l
A

}
_
1
1
_
= (C
l,
D
l,
)
_
1
0
_
+H

_
x
(1)
l,
x
(2)
l,
_
(66)
where H

is dened:
H

=
_
h
(1)

h
(2)

(1)

(2)

_
Using denitions of h
(1)

and h
(2)

from (35) and (36), it is not

difcult to show that the determinant of H

is given by:
det(H

) = (
(2)

(1)

)
This determinant is non-zero if and only if
(1)

=
(2)

, which
is true because

L. Thus, H
1

exists and is given by:

H
1

=
1
(
(2)

(1)

)
_

(2)

h
(2)

(1)

h
(1)

_
However, note from (26) that x
(1)
l,
+x
(2)
l,
= E{Q
l
}, and hence:
E{Q
l
} =
_
1 1

H
1

_
x
(1)
l,
x
(2)
l,
_
= E{Q
l
A

}
_
1 1

H
1

_
1
1
_
(C
l,
D
l,
)
_
1 1

H
1

_
1
0
_
where the last line follows from (66). However, using deni-
tions of h
(1)

and h
(2)

from (35) and (36), it is not difcult to

show that:
_
1 1

H
1

_
1
1
_
=
(
(2)

(1)

) + (h
(1)

h
(2)

)
det(H

)
=
1

Similarly:
_
1 1

H
1

_
1
0
_
=
1

)
Using these identities yields:
E{Q
l
} =
E{Q
l
A

C
l,
D
l,

)
This proves Lemma 3.
APPENDIX D PROOF OF THEOREM 6
A. Time Averages
l
and r
l
From the queueing equation for Y
l
(t) (given in (44)), it is
clear that for all t we have:
1
t
t1

=0

l
()
1
t
t1

=0
R
l
()
Y
l
(t)
t
However, we know from (47) that Y
l
(t) V
l
+ 1 for all t.
Dene r
l
and
l
as time average expectations of the R
l
(t) and

l
(t) processes, respectively. Thus:

l
r
l
for all l L (67)
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 13
B. Computing The Lyapunov Drift
Dene the Lyapunov function L(Q(t)) as in (7). Dene
Y (t) as the vector of ow state (virtual queues), and dene
the combined Lyapunov function (Q(t), Y (t)) as follows:
(Q(t), Y (t)) = L(Q(t)) +
1
2

lL
Y
l
(t)
2
The combined Lyapunov drift (t) is the sum of drift for the
Q(t) and Y (t) terms. The drift for the Q(t) terms is the same
as (22) with the A
l
(t) values replaced by R
l
(t), and the drift
for Y (t) is found from (44) using a standard quadratic drift
argument (see, for example, [6]). Following the framework of
[6] for joint Lyapunov stability and performance optimization,
our goal is to make decisions that minimize a bound for the
expression: (t) V

lL
E{g
l
(
l
(t))}. Omitting the drift
computation details for brevity, we have:
(t) V

lL
E{g
l
(
l
(t))} E{D
Q
(t) +D
Y
(t)}

lL
E{Q
l
(t)}
+

lL
E
_
E
_
R
l
(t)[

Q
S
l
(t) Y
l
(t)] | Q(t), Y (t)
__
+

lL
E{E{Y
l
(t)
l
(t) V g
l
(
l
(t)) | Q(t), Y (t)}} (68)
where

R
S
l
(t)

S
l
R

(t) and Lemma 4 is used, and where

D
Q
(t) and D
Y
(t) are dened:
D
Q
(t)

=
1
2

lL
_
R
l
(t)

R
S
l
(t) 2

R
S
l
(t)
l
(t) +
l
(t)
_
D
Y
(t)

=
1
2

lL
_
R
l
(t)
2
+
l
(t)
2

It is evident that the FLOW-MAXIMAL algorithm observes

the queue values Q(t) and Y (t) every slot t, and chooses
R
l
(t) and
l
(t) to minimize the inside (conditional) expecta-
tions in the last two terms on the right hand side of (68) over
all alternative options. It follows that the above expression is
less than or equal to the corresponding expression when the
decisions R
l
(t) and
l
(t) (for slot t only) are replaced by
R

l
(t) and

l
(t), being any alternative decisions that can be
made on slot t that satisfy the constraints:
0 R

l
(t) A
l
(t) , 0

l
(t) 1 ,

l
(t) I

(69)
Plugging these expressions back into the nal four terms on
the right hand side of (68) and rearranging terms yields:
(t) V

lL
E{g
l
(
l
(t))} E{D
Q
(t) +D
Y
(t)}

lL
E{Q
l
(t)} +

lL

S
l
E{Q
l
(t)R

(t)}

lL
E{Y
l
(t)R

l
(t)} +

lL
E{Y
l
(t)

l
(t)}
V

lL
E{g
l
(

l
(t))} (70)
We emphasize a subtle but important point in the above
expression: The expectations on the right hand side involve
queue backlogs Q
l
(t) and Y
l
(t) that have distributions that
arise from having implemented the FLOW-MAXIMAL al-
gorithm on every timeslot up to time t (so that these are
queue backlogs that arise in the actual algorithm). However,
the R

l
(t) and

l
(t) values are not actually used in the
algorithm implementation, are potentially different from the
FLOW-MAXIMAL decisions, and are considered as being
implemented only on slot t for computation of the expectation.
C. Establishing the Congestion Bound
Consider now the particular decisions R

l
(t) =

l
(t) = 0
for all l L, and note that these indeed satisfy the required
constraints (69). The inequality (70) thus becomes:
(t) V

lL
E{g
l
(
l
(t))} E{D
Q
(t) +D
Y
(t)}

lL
E{Q
l
(t)}
Using this inequality for the drift (t) directly in the Lya-
punov drift Theorem (Theorem 1), for f(t) =

lL
Q
l
(t)
and g(t) = V

lL
g
l
(
l
(t)) +D
Q
(t) +D
Y
(t), yields:
limsup
t
1
t
t1

=0

lL
E{Q
l
()} D
Q
+D
Y
+V limsup
t
1
t
t1

=0

lL
E{g
l
(
l
())}
where D
Q
and D
Y
represent time average expectation of the
D
Q
(t) and D
Y
(t) processes. Using the fact that g
l
(0) = 0,
and that is the largest right derivative of any utility function,
we have g
l
() for all l. Hence, using the fact that steady
state exists, we have:

lL
E{Q
l
} D
Q
+D
Y
+V

lL

l
(71)
where
l
represents the time average expectation of the
l
(t)
process. Using the fact that
l
r
l
(from (67)), it follows
that

lL

l
r
tot
. Further, the D
Q
and D
Y
averages can
be bounded as follows:
D
Y

r
tot
2
[A
max
+ 1]
D
Q

1
2

lL
E
_
A
l
(t)

A
S
l
(t)
_
+
r
tot
2
Using these in (71) proves the average queue bound (49).
D. Establishing the Utility Bound
Without loss of generality, assume that
l
> 0 for all l L
(else, such a link with
l
= 0 can be effectively removed from
the set of links L, as no data ever arrives for transmission
over this link). Dene (r

l
) as the solution to the optimization
problem (40)-(43), so that

lL
g
l
(r

l
) = g

. Note also that

the constraints in (40)-(43) ensure that r

l

l
. For each link
l L, consider the particular decisions for R

l
(t) and

l
(t):

l
(t) = r

l
R

l
(t) =
_
A
l
(t) with probability r

l
/
l
0 otherwise
IEEE TRANSACTIONS ON NETWORKING, VOL. 17, NO. 4, PP. 1146-1159, AUGUST 2009 14
Note that the policy R

l
(t) is randomized, and the randomized
decision is made independently of everything else. These
decisions indeed satisfy the required constraints (69). Hence,
plugging into (70) yields:
(t) V

lL
E{g
l
(
l
(t))} E{D
Q
(t) +D
Y
(t)}

lL
E{Q
l
(t)} +

lL

S
l
E{Q
l
(t)R

(t)}

lL
E{Y
l
(t)R

l
(t)} +

lL
r

l
E{Y
l
(t)} V g

(72)
Note that:
E{Q
l
(t)R

(t) | Z

(t) = m} =
r

(m)

E{Q
l
(t) | Z

(t) = m}
E{Y
l
(t)R

l
(t) | Z
l
(t) = m} =
r

(m)
l
E{Y
l
(t) | Z
l
(t) = m}
Hence, similar to the linear-algebraic derivation in Section V,
we can show that for any time t at which the system is in
steady state (omitting details for brevity):
E{Q
l
(t)R

(t)} E{Q
l
(t)} r

+

C
l,

(73)
E{Y
l
(t)R

l
(t)} E{Y
l
(t)} r

C
l,l

l
+
l
(74)
where

C
l,
is dened:

C
l,

=
E{R
l
(t 1)R

(t)} if

L,
and 0 else. Note that

C
l,
is dened using the actual decision
R
l
(t1) implemented on slot t1 and the alternative decision
R

(t) (not implemented) for slot t, and satises:

C
l,
E{A
l
(t 1)R

(t)} = (r

)E{A
l
(t 1)A

(t)}
Plugging (74) and (73) into (72) yields:
(t) V

lL
E{g
l
(
l
(t))} E{D
Q
(t) +D
Y
(t)}
+

lL

S
l

C
l,

C
l,l

l
+
l
V g

where we have used the fact that

S
l
r

1. Using
the above drift expression in the Lyapunov drift Theorem
(Theorem 1) yields:
limsup
t
1
t
t1

=0

lL
E{g
l
(
l
())} g

F/V (75)
where F is dened as in (51). By concavity of g
l
(r), Jensens
inequality yields:
1
t
t1

=0
E{g
l
(
l
())} g
l
_
1
t
t1

=0
E{
l
()}
_
Taking limits of the above inequality and noting that the time
average expectation
l
exists and satises
l
r
l
for each
l L (by (67)), we have:
limsup
t
1
t
t1

=0

lL
E{g
l
(
l
())}

lL
g
l
(r
l
)
Using this in (75) proves Theorem 6.
REFERENCES
[1] M. J. Neely. Delay analysis for maximal scheduling in wireless networks
with bursty trafc. Proc. IEEE INFOCOM, April 2008.
[2] P. Chaporkar, K. Kar, X. Luo, and S. Sarkar. Throughput and fairness
guarantees through maximal scheduling in wireless networks. IEEE
Trans. on Information Theory, vol. 54, no. 2, pp. 572-594, Feb. 2008.
[3] X. Wu, R. Srikant, and J. R. Perkins. Scheduling efciency of distributed
greedy scheduling algorithms in wireless networks. IEEE Transactions
on Mobile Computing, June 2007.
[4] X. Lin and N. B. Shroff. The impact of imperfect scheduling on cross-
layer rate control in wireless networks. Proc. IEEE INFOCOM, 2005.
[5] L. Tassiulas and A. Ephremides. Stability properties of constrained
queueing systems and scheduling policies for maximum throughput in
multihop radio networks. IEEE Transacations on Automatic Control,
vol. 37, no. 12, pp. 1936-1949, Dec. 1992.
[6] L. Georgiadis, M. J. Neely, and L. Tassiulas. Resource allocation and
cross-layer control in wireless networks. Foundations and Trends in
Networking, vol. 1, no. 1, pp. 1-149, 2006.
[7] D. Shah. Maximal matching scheduling is good enough. Proc. IEEE
Globecom, Dec. 2003.
[8] S. Deb, D. Shah, and S. Shakkottai. Fast matching algorithms for
repetitive optimization: An application to switch scheduling. Proc. of
40th Annual Conference on Information Sciences and Systems (CISS),
Princeton, NJ, March 2006.
[9] J. G. Dai and B. Prabhakar. The throughput of data switches with and
without speedup. Proc. IEEE INFOCOM, 2000.
[10] A. Mekkittikul and N. McKeown. A practical scheduling algorithm
to achieve 100% throughput in input-queued switches. Proc. IEEE
INFOCOM, 1998.
[11] M. J. Neely. Order optimal delay for opportunistic scheduling in
multi-user wireless uplinks and downlinks. Proc. of Allerton Conf. on
Communication, Control, and Computing (invited paper), Sept. 2006.
[12] M. J. Neely, E. Modiano, and Y.-S. Cheng. Logarithmic delay for nn
packet switches under the crossbar constraint. IEEE Transactions on
Networking, vol. 15, no. 3, pp. 657-668, June 2007.
[13] L. Tassiulas. Scheduling and performance limits of networks with
constantly changing topology. IEEE Trans. on Inf. Theory, May 1997.
[14] M. J. Neely, E. Modiano, and C. E Rohrs. Dynamic power allocation and
routing for time varying wireless networks. IEEE Journal on Selected
Areas in Communications, vol. 23, no. 1, pp. 89-103, January 2005.
[15] S. Shakkottai, R. Srikant, and A. Stolyar. Pathwise optimality of the
exponential scheduling rule for wireless channels. Advances in Applied
Probability, vol. 36, no. 4, pp. 1021-1045, Dec. 2004.
[16] N. McKeown, V. Anantharam, and J. Walrand. Achieving 100%
throughput in an input-queued switch. Proc. IEEE INFOCOM, 1996.
[17] E. Leonardi, M. Mellia, F. Neri, and M. Ajmone Marsan. Bounds on
average delays and queue size averages and variances in input-queued
cell-based switches. Proc. IEEE INFOCOM, 2001.
[18] M. J. Neely, E. Modiano, and C. E. Rohrs. Power allocation and routing
in multi-beam satellites with time varying channels. IEEE Transactions
on Networking, vol. 11, no. 1, pp. 138-152, Feb. 2003.
[19] S. Ross. Stochastic Processes. John Wiley & Sons, Inc., New York,
1996.
[20] M. J. Neely, E. Modiano, and C. Li. Fairness and optimal stochastic
control for heterogeneous networks. Proc. IEEE INFOCOM, March
2005.