Scheduling Heterogeneous Real-Time Trafc

over Fading Wireless Channels

I-Hong Hou
CSL and Department of CS
University of Illinois
Urbana, IL 61801, USA
[email protected]
P. R. Kumar
CSL and Department of ECE
University of Illinois
Urbana, IL 61801, USA
[email protected]
AbstractWe develop a general approach for designing
scheduling policies for real-time trafc over wireless chan-
nels. We extend prior work, which characterizes a real-time
ow by its trafc pattern, delay bound, timely-throughput
requirement, and channel reliability, to allow time-varying
channels, allow clients to have different deadlines, and allow
for the optional employment of rate adaptation. Thus, our
model allow the treatment of more realistic fading channels
as well as scenarios with mobile nodes, and the usage of
more general transmission strategies.
We derive a sufcient condition for a scheduling policy
to be feasibility optimal, and thereby establish a class of
feasibility optimal policies. We demonstrate the utility of the
identied class by deriving a feasibility optimal policy for the
scenario with rate adaptation, time-varying channels, and
heterogeneous delay bounds. When rate adaptation is not
available, we also derive a feasibility optimal policy for time-
varying channels. For the scenario where rate adaptation is
not available but clients have different delay bounds, we
describe a heuristic. Simulation results are also presented
which indicate the usefulness of the scheduling policies for
more realistic and complex scenarios.
I. INTRODUCTION
With the wide deployment of Wireless Local Area Net-
works (WLANs) and advances in multimedia technology,
wireless networks are increasingly being used to carry
real-time trafc, such as VoIP and video streaming. These
applications usually specify throughput requirement while
meeting specied delay bounds. We study the problem of
designing scheduling policies for such applications.
While there has been much research on scheduling real-
time trafc over wireline networks, the results are not di-
rectly applicable to wireless networks where channels are
unreliable, with qualities that may be time-varying either
due to fading or node mobility. Also, individual clients
may impose differing delay requirements. These features
present new challenges to the scheduling problems.
We consider the scenario where an Access Point (AP)
is required to serve real-time trafc for a set of clients. A
previous work [10] solves the scheduling problem in a re-
strictive environment and proposes two feasibility optimal
policies. In particular, it assumes a xed transmission rate,
a static channel model, and that all clients in the system
require the same delay bound. We extend this model so
that it can capture the trafc patterns, delay bounds,
timely-throughput bounds, and delivery ratio bounds of
clients, for time-varying wireless channels. We address
scenarios with and without rate adaptation. We establish a
This material is based upon work partially supported by USARO un-
der Contract Nos. W911NF-08-1-0238 and W-911-NF-0710287, AFOSR
under Contract FA9550-09-0121, and NSF under Contract Nos. CNS-07-
21992, ECCS-0701604, CNS-0626584, and CNS-05-19535.
sufcient condition for a scheduling policy to be feasibility
optimal. Based on this we describe a class of policies and
prove that they are all feasibility optimal.
To demonstrate the utility of the class of policies,
we study three particular scenarios of interest. The rst
scenario employs rate adaptation and treats time-varying
channels, as well as allowing different delay bounds for
different clients. The other two scenarios treat the case
where rate adaptation is not available. One scenario con-
siders time-varying channels, while the other considers
the scenario where clients require different delay bounds.
For the former two scenarios, we derive computationally
tractable scheduling policies and prove that they are
feasibility optimal. We also obtain a heuristic for the third
scenario.
We have also tested the derived policies using the
IEEE 802.11 standard in a simulation environment. The
results suggest that the three policies outperform others,
including the policies in [10], and a server-centric policy
that schedules packets randomly. In particular, since the
policies introduced in the previous work fail to provide
satisfactory performance in the environments studied
here, this suggests that neglecting the facts that the sys-
tem can apply rate adaptation, that wireless channels are
time-varying, and the possibility that clients may require
different delay bounds, can result in malperformance of
the derived policies.
Section II reviews some of the related work. Section
III describes the extension of the model in [10]. Section
IV discusses some useful observations for scheduling and
reviews policies proposed in [10]. In Section V, we study
an extension for time-varying channels. In Section VI,
we derive a general class of policies that are feasibility
optimal. Based on this class, we obtain scheduling policies
in Sections VII and VIII, and a heuristic in Section IX, for
different scenarios. In Section X, we discuss implementa-
tion issues and simulation results. Section XI concludes
the paper.
II. RELATED WORK
The problem of providing QoS over unreliable wireless
channels has received growing interest in recent years.
Tassiulas and Ephremides [18] have considered the prob-
lem in a single-hop network by assuming ON/OFF chan-
nels and derived a throughput-optimal policy. Though the
policy is unaware of packet delay, Neely [15] has shown
that average packet delay is constant regardless of the net-
work size. Andrews et al [1] have proposed another policy
that aims to improve packet delay. They have proved
that their policy is also throughput optimal but offer no
2
theoretical bound on packet delays. Liu, Wang, and Gi-
annakis [14] have used a cross-layer approach to provide
differentiated service for a variety of classes of clients.
Grilo, Macedo, and Nunes [8] have proposed a resource-
allocation algorithm based on the expected transmission
time of each packet. Since the expected transmission
time may not be an accurate indication of the actual
transmission time, their work cannot provide provable
delay guarantees. Raghunathan et al [16] and Shakkottai
and Srikant [17] have both approached this problem by
analytically demonstrating algorithms to minimize the
total number of expired packets in the system. Their
results, however, cannot provide differentiated service to
different clients. Hou, Borkar, and Kumar [9] have studied
the problem of providing QoS based on delay bounds and
delivery ratio requirements, and proposed two optimal
policies under some restrictive assumptions. Their work
has been further extended to deal with variable-bit-rate
trafc [10]. In this paper, we extend this work to more
realistic scenarios, including rate adaptation, time-varying
channels and heterogeneous delay bounds among clients.
Fattah and Leung [6] and Cao and Li [3] have surveyed
other existing scheduling policies for providing QoS.
III. SYSTEM MODEL
We begin by extending the model proposed in [10],
which only considers a static channel condition and xed
delay bounds for all clients, to account for network be-
havior and application requirements for providing QoS in
wireless systems.
Consider a wireless system with N clients,
1, 2, . . . , N, and one access point (AP). Packets
for clients arrive at the AP. Time is slotted with slots
t 0, 1, 2, . . . . Time slots are further grouped into
periods [kT, (k + 1)T) with period length T. Packets
arrive at the AP at the beginning of each period, at time
slots 0, T, 2T, . . . , probabilistically, with no more than
one packet per client. We model the packet arrivals as a
stationary, irreducible Markov process with nite state.
The average probability that packets arrive for subset
S of clients is R(S). Packet arrivals can be dependent
between clients, and packet arrivals in a period can
depend on other periods.
Each client n species a delay requirement
n
, with

n
T. If the packet for client n is not delivered by the

n
th
time slot of the period, the packet expires and is
discarded. This scheme applies naturally to a wide range
of server-centric wireless communication technologies,
such as IEEE 802.11 Point Coordination Function (PCF),
WiMax, and Bluetooth.
We consider an unreliable, heterogeneous, and time-
varying channel model. We model the channel condition
as a stationary, irreducible Markov process with a nite set
of channel states (. The average probability that channel
state c occurs is f
c
and the channel state remains constant
within each period. We consider the system both with
rate adaptation and without. When rate adaptation is not
available, that is, when all packets are transmitted at a
xed rate, the AP can make exactly one transmission in
each time slot. Under channel state c, the link reliability
between the AP and client n is p
c,n
, so that a packet
transmitted by the AP for client n is delivered with
probability p
c,n
. On the other hand, when the system uses
rate adaptation, the channel states describe the maximal
rates that can be supported between the AP and clients,
which in turn decide the service times for transmissions.
Under channel state c, it takes s
c,n
time slots to make an
error free transmission to client n.
The channel state and the packet arrivals in a period
are assumed to be independent of each other. We also
assume that the AP has knowledge of channel state, as
well as whether a transmission is successful, for example,
through ACKs, in which case p
c,n
is the probability that
the AP receives an ACK after making a transmission.
Each client n requires a timely-throughput of at least
q
n
packets per period. Since, on average, there are

S:nS
R(S) packets for client n per period, this timely-
throughput bound can also be interpreted as a delivery
ratio requirement of
qn

S:nS
R(S)
.
Denition 1: A set of clients, 1, 2, . . . , N is fullled
under a scheduling policy , if for every > 0,
Prob
d
n
(t)
t/T
> q
n
, for every n 1, as t ,
where d
n
(t) is the number of packets delivered to client
n up to time t.
IV. SCHEDULING POLICIES
Since the overall system can be viewed as a controlled
Markov chain, we have:
Lemma 1: For any set of clients that can be fullled,
there exists a stationary randomized policy that fullls the
clients, which uses a probability distribution based only
on the channel state, the set of undelivered packets, and
the number of time slots remaining in the system (and
not any events depending on past periods), according
to which it randomly chooses an undelivered packet to
transmit, or stays idle.
Since the computational overhead for some complex
policies may be too high for real-time applications, we
consider the limited set of priority-based policies, which
require computation only at the beginning of each period:
Denition 2: A priority-based policy is a scheduling pol-
icy which assigns priorities to some of the clients, based
on past history and current state of the system, at the
beginning of each period. During the period, a packet for
a client is transmitted only after all packets for clients with
higher priorities have been delivered. Packets for clients
which do not receive a priority are never transmitted. A
stationary randomized priority-based policy is one which
chooses the priority order randomly according to a prob-
ability distribution that depends only on the channel state
and packet arrivals at the beginning of each period. We
denote by P and P
rand
the sets of priority-based policies
and stationary randomized priority-based policies.
Denition 3: A set of clients is feasible in the set P (or
P
rand
) if there exists some scheduling policy in P (or
P
rand
) that fullls it.
Similar to Lemma 1, if [q
n
] is feasible in the set P, it is
also feasible in the set P
rand
.
Denition 4: We call the region in the N-space formed
by vectors [q
n
] for which the clients are feasible in P (or
all policies), as the feasible region under P (or all policies).
Lemma 2: The feasible region under the class of all
policies, or P, are both convex sets.
Proof: Let [q
n
] and [q

n
] be two vectors in the feasible
region under P, and thus also feasible in P
rand
. Let
and

be policies in P
rand
that fulll the two vectors,
respectively. Then, the policy in P that randomly picks one
3
of the two policies, with being chosen with probability
, at the beginning of each period, fullls the vector
[q
n
+(1 )q

n
]. Further, since q
n
and q

n
are both larger
than 0 for each n, q
n
+(1)q

n
> 0 for all n. Thus, the
vector [q
n
+ (1 )q

n
] also falls in the feasible region
under P. A similar proof holds for the class of all policies.
Note that if [q
n
] is feasible in P, then so is [q

n
], where
0 < q

n
q
n
.
Denition 5: [q
n
] is strictly feasible in P (or the class of
all policies) if there exists some (0, 1) such that [q
n
/]
is feasible in P (or the class of all policies).
1
Denition 6: A scheduling policy is feasibility optimal
among P (or the class of all policies) if it fullls every set
of clients that is strictly feasible in P (or the class of all
policies).
In the rest of the paper, unless otherwise specied, the
default is the set of all policies.
A. The Static Channel Case
In previous work [10], the problem of admission con-
trol and feasibility optimal scheduling has been addressed
for the case where the channel state is static, and all
clients require the same delay bounds, i.e. [([ = 1 and

n
. In the special case, we will use p
n
instead of p
c,n
since the channel state is static, and instead of
n
.
Two largest debt rst scheduling polices were proved
to be feasibility optimal, where the AP, based on the
past history, calculates a debt for each client. In each
period, the AP sorts all clients according to their debts,
and schedules a packet for client n only after all packets
for clients with larger debts have been delivered. The
rst policy, the largest time-based debt rst policy, uses
the time-based debt for client n at time slot t, dened
as
t
T
w
n
minus the number of time slots that the AP has
spent on transmitting packets for client n up to time slot
t. The other policy, the largest weighted-delivery debt rst
policy, uses the weighted-delivery debt for client n at time
slot t, dened as
t
T
qndn(t)
pn
, where d
n
(t) is the number of
delivered packets for client n up to time slot t.
As for admission control, the following lemma was
proved in [10]:
Lemma 3: A set of clients is fullled if and only if
the long-term average number of time slots that the AP
spends on transmitting packets for client n per period is
at least w
n
=
qn
pn
for each n.
Further, since expired packets are dropped, the number
of packets in the system is bounded. Thus, there may
be some time slots where the AP may have delivered
all packets in the system, and is therefore forced to stay
idle. For any subset S of 1, 2, . . . , N, dene I
S
to be
the minimum number of time slots that the AP is idle
in a period for any scheduling policy, given that the AP
can only transmit packets for the subset S of clients. A
necessary and sufcient condition for strict feasibility is
proved:
Theorem 1: A set of clients is strictly feasible if and only
if

nS
w
n
< T E[I
S
], for all S 1, 2 . . . , N.
1
Equivalently, [qn] is an interior point of the feasible region under P
(or the class of all policies).
V. TIME-VARYING CHANNELS
We now discuss how to extend the aforementioned
policies to provide QoS for time-varying channels. One
intuitive approach is to decouple the channel states. The
AP assigns a timely-throughput requirement q
c,n
for each
channel state c and client n, with

cC
f
c
q
c,n
q
n
. Also,
for each channel state c, the assigned throughput require-
ments must be strictly feasible under that channel state,
that is,

nS
qc,n
pc,n
< T E[I
c,S
] for all S 1, 2, . . . , N ,
where I
c,S
is the minimal number of time slots that the AP
is forced to stay idle in a period under channel state c for
any scheduling policy, given that the AP only transmits
packets for the subset S of clients. More formally, we
therefore seek a matrix Q = [q
c,n
] that solves the following
linear programming problem:
Max

N
n=1

cC
f
c
q
c,n
s.t.

cC
f
c
q
c,n
q
n
, n

nS
q
c,n
p
c,n
< T E[I
c,S
], c, S 1, 2, , N.
After obtaining the matrix Q, we can modify the two
largest debt rst policies to deal with time-varying chan-
nel conditions. Let s
c
(t) be the number of time slots up to
time slot t that the channel state has been c, and assume
that the channel state at time slot t is c. In the largest
time-based debt rst policy, we dene the time-based debt
for client n under channel state c as
sc(t)
T
qc,n
pc,n
minus the
number of time slots that the AP has spent on transmitting
packets for client n under channel state c up to time
slot t. In the largest weighted-delivery debt rst policy,
we dene the weighted-delivery debt for client n under
channel state c as
sc(t)
T
qc,ndc,n(t)
pc,n
, where d
c,n
(t) is the
number of delivered packets for client n under channel
state c. Obviously, these two modied largest debt rst
policies are feasibility optimal.
While this extension offers feasibility optimality, the
above linear program involves exponentially many con-
straints. Further, it also requires the knowledge of the
distribution [p
c,n
] of channel states. In many scenarios,
such as those with mobile nodes, this knowledge may not
be available. This motivates us, in the following sections,
to describe a more general class of feasibility optimality
policies, and derive an on-line scheduling policy that is
feasibility optimal for the time-varying channel condi-
tions.
VI. A SUFFICIENT CONDITION FOR FEASIBILITY
OPTIMALITY
We now describe a more general class of policies that
is feasibility optimal. We start by extending the concept
of debt.
Denition 7: A variable r
n
(k), whose value is deter-
mined by the past history of the client n up to the k
th
period, or time slot kT, is called a pseudo-debt if:
1) r
n
(0) = 0, for all n.
2) At the beginning of each period, r
n
(k) increases by a
constant strictly positive number z
n
= z
n
(q
n
), which
is an increasing linear function of q
n
.
3) r
n
(k+1) = r
n
(k) +z
n
(q
n
)
n
(k), where
n
(k) is a
non-negative and bounded random variable whose
value is determined by the behavior of client n.
4
Further,
n
(k) = 0 if the AP does not transmit any
packet for client n.
4) The set of clients is fullled if and only if
Prob
rn(k)
k
< 1, as k , for all n and
all > 0.
In the following example, we illustrate that both
the time-based debt and the weighted-delivery debt are
pseudo-debts under a static channel model.
Example 1: At the beginning of each period, the time-
based debt r
(1)
n
(k) increases by w
n
=
qn
pn
, and decreases
by the number of time slots that the AP has transmitted
packets for client n during the period. Lemma 3 shows
that condition (4) is satised.
Similarly, r
(2)
n
(k), the weighted-delivery debt is also a
special case. It increases by
qn
pn
at the beginning of each
period, and decreases by
1
pn
if a packet is delivered for
client n during that period, and 0 otherwise. It satises
condition (4) by denition. 2
We can also dene the feasible region for debt in P (or
in the set of all policies) as the set of [z
n
] such that
the corresponding [q
n
] is feasible in P (or in the set of
all policies). Since z
n
is a linear function of q
n
and the
feasible region for [q
n
] is a convex set (Lemma 2), the
feasible region for [z
n
] is also a convex set.
Using the concept of pseudo-debt, we prove a sufcient
condition for feasibility optimality. The proof resembles
one used by Neely [15], though in a different context,
and is based on:
Theorem 2 (Lyapunov Drift Theorem): Let L(t) be a
non-negative Lyapunov function. Suppose there exists
some constant B > 0 and non-negative function f(t)
adapted to the past history of the system such that:
EL(t + 1) L(t)[history up to time t B f(t),
for all t, then: limsup
t
1
t

t
i=0
Ef(i) B/. 2
Theorem 3: Let r
n
(k) be a pseudo-debt.
1) A policy that maximizes the payoff function
N

n=1
Er
n
(k)
+

n
(k)[c
k
, S
k
, [r
m
(k)] (1)
at the beginning of each period is feasibility opti-
mal, where c
k
denotes the channel state in the k
th
period, and S
k
is the subset of clients whose packets
arrive at the AP at the beginning of the k
th
period.
2) A priority-based policy that maximizes (1) over all
policies in P is feasibility optimal in P.
Proof: We present the proof for P only. A similar proof
works for the class of all policies too. Dene L(k) =
1
2

N
n=1
r
n
(k)
2
. Since r
n
(k + 1) = r
n
(k) + z
n

n
(k),
(L(k)) := EL(k + 1) L(k)[[r
m
(k)]
=E
1
2
N

n=1
r
n
(k + 1)
2

1
2
N

n=1
r
n
(k)
2
[[r
m
(k)]
=E
N

n=1
r
n
(k)[z
n

n
(k)] +
1
2
N

n=1
[z
n

n
(k)]
2
[[r
m
(k)].
Dene B(k) := E
1
2

N
n=1
[z
n

n
(k)]
2
[[r
m
(k)]. Then
B(k) B, for all k, for some B. Hence for any policy in
P:
(L(k)) E
N

n=1
r
n
(k)[z
n

n
(k)][[r
m
(k)] + B. (2)
Suppose [q
n
] is strictly feasible in P. The vector [z
n
] is
thus an interior point of the feasible region (for debt)
under P, and there therefore exists some (0, 1) such
that [z
n
/] is also in the feasible region under P. Let
z
min
= minz
1
, z
2
, . . . , z
N
. The N-dimensional vector
[z
min
] whose elements are all z
min
, falls in the feasible re-
gion under P. Since the feasible region under P is a convex
set, the vector [z
n
/] +(1)[z
min
] = [z
n
+(1)z
min
]
is also in the feasible region under P.
By Lemma 1, there exists a stationary randomized
policy

in P that fullls the set of clients with timely-

throughput bounds for the vector [z
n
+ (1 )z
min
]. Let

n
(k) be the decrease in the pseudo-debt for client n
under

during the period. Then, we have:

E

n
(k)[[r
m
(k)] = EE

n
(k)[c
k
, S
k
, [r
m
(k)]
z
n
+ (1 )z
min
.
Above, the outer expectation in the RHS is taken over
channel states and the vectors of packet arrivals.
Let be a policy that maximizes the payoff function
(1), for all k, among all policies in P. Then dening
n
(k)
and r
n
(k) as the decrease resulting from policy and the
pseudo-debt, we have:

N
n=1
Er
n
(k)
+

n
(k)[c
k
, S
k
, [r
m
(k)]

N
n=1
Er
n
(k)
+

n
(k)[c
k
, S
k
, [r
m
(k)].
We can assume without loss of generality that the policy
does not work on any client n with r
n
(k) 0, that is,

n
(k) = 0 if r
n
(k) 0.
2
From (2), we obtain:
(L(k)) E

N
n=1
r
n
(k)
+
[z
n

n
(k)][[r
m
(k)] + B
E

N
n=1
r
n
(k)
+
[z
n

n
(k)][[r
m
(k)] + B

N
n=1
r
n
(k)
+
(1 )z
min
+ B.
Let := (1 )z
min
. By Theorem 2,
limsup
k
1
k

k
i=0
E

N
n=1
r
n
(k)
+
B/. (3)
Finally, since z
n
is a constant and
n
(k) is a bounded
function, [r
n
(k + 1) r
n
(k)[ is bounded, which implies
that [

N
n=1
r
n
(k + 1)
+

N
n=1
r
n
(k)
+
[ is also bounded
for all k. Thus, (3) implies that
1
k
E

N
n=1
r
n
(k)
+
0
as k , as shown in Lemma 4 below. This shows that
rn(k)
+
k
converges to 0 in probability for all n. Hence, is
feasibility optimal in P.
Lemma 4: Let f(t) be a non-negative function such that
[f(t + 1) f(t)[ M, for some M > 0, for all t. If
limsup
t
1
t

t
i=0
f(i) B/, then lim
t
1
t
f(t) = 0.
Proof: We prove by contradiction. Suppose
limsup
t
1
t
f(t) > , for some > 0. Thus,
f(t) > t innitely often. Suppose f(t) > t for
some t. Since [f(t) f(t 1)[ < M, we have
f(t 1) > t M. Similarly, f(t 2) > t 2M,
f(t3) > t3M, . . . , f(tt/M|) > tt/M|M 0.
2
Since a policy cannot lose its feasibility optimality by doing more
work, this assumption is not restrictive.
5
Summing over these terms gives:

t
i=tt/M
f(i) >
tt/M
2
, and thus,

t
i=0
1
t
f(i) >
t/M
2
. Since f(t) > t
innitely often, limsup
t

t
i=0
1
t
f(i) = , which is a
contradiction.
Theorem 3 suggests a more general procedure to de-
sign feasibility optimal scheduling policies. To design a
scheduling policy in a particular scenario, we need to
choose an appropriate pseudo-debt and obtain a policy
to maximize the payoff function. Maximizing the payoff
function is, however, in general, difcult. Nevertheless, in
some special cases, evaluating the payoff function gives
us simple feasibility optimal policies, or, at least, some
insights into designing a reasonable heuristic, as long
as we choose the correct pseudo-debt. In the following
sections, we demonstrate the utility of this approach.
VII. SCHEDULING POLICY WITH RATE ADAPTATION
We now propose a feasibility optimal scheduling policy
when rate adaptation is employed. Channel qualities can
be time-varying and clients may have different deadlines.
To derive the scheduling policy, we dene the delivery
debt r
(3)
n
(k) := q
n
kd
n
(kT), where d
n
(t) is the number of
delivered packets for client n up to time slot t. Thus, z
n
:=
q
n
, while
n
(k) = 1 if a packet for client n is delivered in
the period, and
n
(k) = 0 otherwise.
Suppose at the beginning of period k, the delivery debt
vector is [r
(3)
n
(k)], the channel state is c, and the set of
arrived packets is S. The transmission time for client n is
s
c,n
time slots, and client n stipulates a delay bound of

n
. Since transmissions are assumed to be error-free when
rate adaptation is applied, the scheduling policy consists
of nding an ordered subset S

= m
1
, m
2
, . . . , m
N
of
S such that

l
n=1
s
c,n

l
, for all 1 l m
N
. That is,
when clients are scheduled according to the ordering, no
packets for clients in S

would miss their respective delay

bounds. By Theorem 3, a policy using an ordered set S

that maximizes

nS
r
(3)
n
(k) with the above constraint
is feasibility optimal. This is a variation of the knapsack
problem. When S

is selected, reordering clients in S

in
an earliest-deadline-rst fashion also allows all packets
to meet their respective delay bounds. Based on this
observation, we derive the feasibility optimal scheduling
algorithm, the Modied Knapsack Algorithm. Let M[n, t]
be the maximum debt a policy can collect if only clients 1
through n can be scheduled and all transmissions need to
complete before time slot t. Thus, max
S

nS
r
(3)
n
(k) =
M[N, T]. Also, iteratively:
M[n, t] =

M[n, t 1] if t >
n
,
maxM[n 1, t],
r
(3)
n
(k) + M[n 1, t s
c,n
] otherwise,
where M[n 1, t] is the maximum debt can be collected
when client n is not scheduled, and r
(3)
n
(k) +M[n1, t
s
c,n
] is that when client n is scheduled. The complexity of
this algorithm is O(N), and it is thus reasonably efcient.
VIII. COMPUTATIONALLY TRACTABLE SCHEDULING FOR
TIME-VARYING CHANNELS
We now consider the case when rate adaptation is
not available, and propose a scheduling policy for time-
varying channels and homogeneous delay bounds. We
show that the policy is feasibility optimal among all
Algorithm 1 Modied Knapsack Policy
1: for n = 1 to N do
2: r
(3)
n
(k) = q
n
k d
n
(kT)
3: Sort clients such that
1

2

N
4: S

[0, 0] =
5: M[0, 0] = 0
6: for n = 1 to N do
7: for t = 1 to T do
8: if t >
n
then
9: M[n, t] = M[n, t 1]
10: S

[n, t] = S

[n, t 1]
11: else if client n has a packet AND
r
(3)
n
(k) + M[n 1, t s
c,n
] > M[n 1, t] then
12: M[n, t] = r
(3)
n
(k) + M[n 1, t s
c,n
]
13: S

[n, t] = S

[n 1, t s
c,n
] + n
14: else
15: M[n, t] = M[n 1, t]
16: S

[n, t] = S

[n 1, t]
17: schedule according to S

[N, T]
priority-based policies. We use the delivery debt, r
(3)
n
(k),
of Section VII.
Suppose at the beginning of a period, the delivery
debt vector is [r
(3)
n
(k)], the channel state is c, and the
set of arrived packets is S. We wish to nd the priority
ordering that maximizes the payoff function
tot
(k) =

N
n=1
r
(3)
n
(k)
+
E
n
(k), where in the expectation we
suppose that the channel state c and the set of arrival
packets S are both xed. Obviously, transmitting a packet
from a client n with r
(3)
n
(k) 0 will not increase the value
of
tot
(k). Thus, we do not give priorities to clients with
non-positive delivery debts. For ease of the remaining
discussion, we further assume r
(3)
n
(k) > 0 for all n.
Consider two orderings, A and B: In A, the priority
order is 1, 2, . . . , N, while, in B, the priority order is
1, 2, . . . , m 1, m + 1, m, m + 2, m + 3, . . . , N. Let the
values of the payoff functions be
A
tot
and
B
tot
. Since
clients 1 through m 1 have the same priorities in
both orderings and their priorities are higher than the
remaining clients, the values of E
n
(k), 1 n m1
are the same for both orderings. On the other hand,
clients m + 2 through N also have the same priorities
in both orderings and they can be scheduled only after
the packets for clients 1 through m+1 are delivered. The
probabilities of packet deliveries for these clients are the
same under the two orderings. Thus, to compare the two
orderings, one only needs to evaluate the probabilities of
packet delivery for client m and m+1. We further notice
that the probabilities that packets for both clients m and
m+1 are delivered are also the same for both orderings.
With e
n
the event that the packet for client n is delivered,

A
tot

B
tot
=r
(3)
m
(k)Probe
m
e
m+1
[ordering A
r
(3)
m+1
(k)Probe
m+1
e
m
[ordering B.
Suppose that there are

time slots left when all

packets from client 1 through m1 have been delivered.
The probability distribution of

is the same under both

6
orderings. Since the channel reliability is p
c,n
,

A
tot

B
tot
=r
(3)
m
(k)E

t=1
p
c,m
(1 p
c,m
)
t1
(1 p
c,m+1
)

r
(3)
m+1
(k)E

t=1
p
c,m+1
(1 p
c,m+1
)
t1
(1 p
c,m
)

=[r
(3)
m
(k)p
c,m
r
(3)
m+1
(k)p
c,m+1
]
E

1
t=0
(1 p
c,m
)
t
(1 p
c,m+1
)

t1
.
Thus,
A
tot

B
tot
if r
(3)
m
(k)p
c,m
r
(3)
m+1
(k)p
c,m+1
. This
leads us to obtain the Joint Debt-Channel Policy. The
computation time is only O(N log N).
Algorithm 2 Joint Debt-Channel Policy
1: for n = 1 to N do
2: r
(3)
n
(k) = q
n
k d
n
(kT), for all n
3: Sort clients with a packet arrival such that
r
(3)
1
(k)p
c,1
r
(3)
2
(k)p
c,2
r
(3)
N0
(k)p
c,N0
>
0 r
(3)
N0+1
(k)p
c,N0+1
. . .
4: Transmit packets for clients 1 through N
0
by the
ordering
Theorem 4: The joint debt-channel policy is feasibility
optimal among all priority-based policies.
Proof: Let be the joint debt-channel policy and

any priority-based policy. Suppose the priorities assigned

by the policies are
1
,
2
, . . . ,
m
, and

1
,

2
, . . . ,

m
. We
modify

as follows:
1) Delete any element in

m
with r
(3)

n
(k) 0.
2) For any client n with r
(3)
n
(k) > 0 that is not in

m
, append it at the end of the ordering.
3) If

m
is still different from
1

m
, there ex-
ists some n such that r
(3)

n
(k)p
c,

n
< r
(3)

n+1
(k)p
c,

n+1
.
Swap

n
and

n+1
.
4) Repeat Step 3 until the two orderings are the same.
Steps 1 and 2 will not decrease the value of the payoff
function. As derived above, Step 3 does not decrease the
value of the payoff function, either. Thus, maximizes
the payoff function and is feasibility optimal in P.
IX. A HEURISTIC FOR HETEROGENEOUS DELAY BOUNDS
We now describe a heuristic for packet scheduling, for
the case where each channel state is static and trans-
mission rate is xed, but clients require different delay
bounds. We use p
n
to represent channel reliability.
We will use the time-based debt, r
(1)
n
(k), as
discussed in Example 1. The payoff function is
E

N
n=1
r
(1)
n
(k)
+

n
(k).
Suppose, without loss of generality, that at the begin-
ning of a period, packets for clients 1, 2, . . . , N
0
arrive.
We further assume that
1

2

N0
. Let
n
be the number of transmissions the AP needs to make
for client n for success. While
n
is a random variable
that cannot be foretold, we examine how to maximize

N0
n=1
r
(1)
n
(k)
+

n
(k) if we knew
n
.
We solve this by proceeding backwards in time. Dur-
ing time slots [
N01
+ 1,
N0
], all packets except the
one for client N
0
have expired, and we can only make
transmissions for client N
0
during these time slots. Thus,
it does not make sense to schedule client N
0
for more
than
N01
N0
:=
N0
(
N0

N01
) transmissions before
time slot
N01
. Next, in the time slots between [
N02
+
1,
N01
], only clients N
0
1 and N
0
can be scheduled. An
obvious choice is to schedule the client with larger debt
rst, with the restriction that it is not scheduled for more
than
N01
n
time slots, and to then schedule the other
client. (For simplicity, we let
N01
N01
:=
N01
.) We can
further obtain the remaining transmissions allowed for
client n before time slot
N02
, which we call
N02
n
, as

N01
n
minus the number of transmissions scheduled for
client n during time slots [
N02
+1,
N01
]. Transmissions
of the remaining time slots are scheduled similarly.
While it is impossible to know the exact value of
n
in
advance, we can estimate it. One estimate is its expected
value,
1
pn
. However, this estimate does not consider the
timely-throughput requirements. If a client has signi-
cantly larger debt than others, a reasonably good policy
would allocate enough time slots so that the probability
of packet delivery for the client in this period is at least
its delivery ratio bound,
qn

nS
R(S)
, given that a packet
for client n arrived. So we estimate
n
by the number
of transmissions that we need to allocate for client n so
that it can achieve its delivery ratio bound. Since the
channel reliability for client n is p
n
, this estimate
n
is
log
1pn
(1
qn

nS
R(S)
)|. We thus derive the Adaptive-
Allocation Policy shown in Algorithm 3.
As a nal remark, note that in all the three policies dis-
cussed in this paper, we do not schedule transmissions for
clients with non-positive debts. This restriction improves
the performance for clients with non-real time trafc. In
practice, it is possible that clients with real-time trafc
and clients with non-real time trafc coexist. Thus, it is
important not to allocate too much of the resource to real-
time clients and starve those with non-real time trafc.
Algorithm 3 Adaptive-Allocation Policy
1: for n = 1 to N do
2: r
(1)
n
(k) = time-based debt
3:
n
= log
1pn
(1
qn

nS
R(S)
)|
4: Sort clients so that packets for clients 1 N
0
arrive
and r
(1)
1
(k) r
(1)
2
(k) r
(1)
N0
(k)
5: alloc n 1vector
6: for t = T to 1 do
7: n 1
8: while (
n
> t or
n
0) and n N
0
do
9: n n + 1
10: if r
(1)
n
(k) > 0 then
11: alloc[t] n
12: else
13: alloc[t] N
0
+ 1
14: if n N
0
then
15:
n

n
1
16: for each time slot t do
17: if alloc[t] N
0
and the packet for client alloc[t] has
not been delivered then
18: transmit the packet for client alloc[t]
19: else
20: transmit the packet with the largest positive time-
based debt
7
TABLE I: MPEG Trafc Pattern
Activity Great High Regular
Data rate 501597 392237 366587
Arrival probability 1 0.8 0.75
X. SIMULATION RESULTS
We have implemented the scheduling policies discussed
in previous sections by using the IEEE 802.11 PCF stan-
dard in the ns-2 simulator. We present the simulation
results for the scenario with time-varying channels, and
with clients requiring different delay bounds. In each
scenario, we compare our policies against the two largest
debt rst policies of [10], and a policy that assigns
priorities to clients randomly, random. IEEE 802.11e, an
enhancement to 802.11 for QoS, allows clients with real-
time trafc to use smaller contention window and inter
frame space to obtain priorities over clients with non-
real time trafc. However, clients with real-time trafc
have to compete with each other in a random access
manner with equal channel access probabilities, without
any QoS based preference or discrimination. Further, the
inter frame space and contention window size are smaller
in PCF than in 802.11e. Thus, the random policy can be
viewed as an improved version of 802.11e. Similar to the
previous work, we conduct two sets of simulations for
each scenario, one with clients carrying VoIP trafc, and
one with clients carrying video streaming trafc. The ma-
jor difference between the two settings lies in their trafc
patterns. Many VoIP codecs generate packets periodically.
Thus, future packet arrivals can be easily predicted and
may be dependent among different clients. For example, if
two clients generate packets at the same rate, then either
all or none of their packets arrive simultaneously. On the
other hand, video streaming technology, such as MPEG,
may generate trafc with variable-bit-rate (VBR). Thus,
packets arrive at the AP probabilistically, with probability
depending on the context of the current frame, and
arrivals are independent among different clients.
For the VoIP trafc, we follow the standards of the
ITU-T G.729.1 [12] and G.711 [11] codecs. Both codecs
generate trafc periodically. G.729.1 generates trafc with
bit rates 8 32 kbits/s, while G.711 generates trafc at
a higher rate of 64 kbits/s. We assume the period length,
T, is 20 ms, and the payload size of a packet is 160 Bytes.
The codecs generate one packet every several periods;
with the duration between packet arrivals depending on
the bit rate used.
We use MPEG for the video streaming setting. MPEG
VBR trafc is usually modeled as a Markov chain consist-
ing of three activity states [13] [5]. Each state generates
trafc probabilistically at different mean rates, with the
state being determined by the current frame of the video.
The statistical mean rates in each state are those obtained
in an experimental study [5]. We use them in setting the
trafc patterns of MPEG trafc. We assume the period
length to be 6 ms and the payload size of a packet to
be 1500 Bytes. Table I shows the statistical results of the
experimental study [5], where we also present them in
terms of the packet arrival probability of our setting. In
Table I, Data rate is measured in bits/GoP, where 1
GoP= 240 ms.
We simulate 20 runs for each setting, each run lasting
one minute in simulated time. All results shown are
averaged over the 20 runs. A natural performance metric
for a client is the delivery debt, r
(3)
n
(k). The performance
of the system is measured by the sum of the positive
delivery debts of the clients, that is,

N
n=1
r
(3)
n
(k)
+
, the
total delivery debt. In addition to evaluating how well the
tested policies serve clients with real-time trafc, we also
wish to know whether the policies starve those with non-
real time trafc. Hence we add a client with saturated
non-real time trafc in all simulations. Packets for the
non-real time client are scheduled in all time slots that
are left idle otherwise. We measure the throughput of the
client with non-real time trafc by the average number of
packets delivered.
A. Rate Adaptation
We present the simulation results under the scenario
where rate adaptation is applied, channels are time-
varying, and clients may require different delay bounds.
We rst show the results for VoIP trafc. We use IEEE
802.11b as the MAC protocol, which can provide a max-
imum data rate of 11 Mb/s. We assume that the chan-
nel capacity of each client alternates between 11 Mb/s
and 5.5 Mb/s. Simulation results suggest that the times
needed for a transmission, including all MAC overheads
such as the time for waiting an ACK, are around 480 s
and 610 s for the two transmission rates, respectively.
Ideally, the length of a time slot should be a common
divisor of the transmission times needed under the two
used data rates. We approximate this value by 160 s.
Thus, transmitting a packet requires 3 time slots when
using 11 Mb/s and 4 time slots when using 5.5 Mb/s.
Further, a period consists of 125 time slots.
There are two groups of clients, A and B. Clients in
group A generate one packet every three periods, or
at rate 21.3 kbits/s, and require 90% of each of the
clients packets to be delivered, or a timely-throughput
requirement of 19.2 kbits/s. Clients in group B generate
one packet every two periods at rate 32 kbits/s, and
require 70% of each of the clients packets to be deliv-
ered, corresponding to a timely-throughput requirement
of 22.4 kbits/s. The two groups can be further divided
into subgroups, A
1
, A
2
, A
3
, B
1
, and B
2
, each with 22
clients. Clients in subgroup A
i
generate packets at periods
[i, i + 3, i + 6, . . . ], and clients in subgroup B
i
generate
packets at periods [i, i + 2, i + 4, . . . ]. Finally, clients in
group A require a delay bound equal to the period length,
or 125 time slots, while clients in group B require a delay
bound equal to two-third of the period length, or 83 time
slots.
Simulation results are shown in Figure 1. The modied
knapsack policy incurs the least total delivery debt among
all evaluated policies. This is because all the other three
policies neglect the time-varying channels with different
data rates and the heterogeneous delay bounds. Fur-
ther, by only scheduling those clients with positive deliv-
ery debts, the modied knapsack policy achieves higher
throughput for the non-real time client than both the
policies proposed in [10]. The random policy results in the
highest throughput for the non-real time client. However,
this is because it sacrices the real-time clients. In fact,
its total delivery debt is more than 300 times larger
than the total delivery debt of the modied backpack
policy. This huge difference suggests that the random
policy, and therefore also 802.11e, are not adequate for
providing QoS when multiple clients with real-time trafc
are present.
8
Fig. 1: Performance for VoIP trafc with rate adaptation.
Fig. 2: Performance for MPEG trafc with rate adaptation.
Next we consider the scenario with MPEG trafc. Since
video streaming requires much higher bandwidth than
VoIP, we use 802.11a as the underlying MAC, which can
support up to 54 Mb/s. We assume that channel capacity
for each client alternates between 54 Mb/s and 24 Mb/s.
The transmission times for a data-ACK handshake require
660 s with 54 Mb/s data rate, and 940 s with 24 Mb/s.
The length of a time slot is 60 s. Thus, the transmission
times for the two data rates are 11 time slots and 16 time
slots, respectively. Further, a period consists of 100 time
slots.
We again assume there are two groups of clients. Clients
in group A generate packets according to Table I, and
clients in group B are assumed to offer only lower quality
video by generating packets only 80% as often as those in
group A, in each of the three states. We assume clients in
group A require 90% delivery ratios, and clients in group
B require 60% delivery ratios. Since the length of a period
for MPEG is very small, it is less meaningful to discuss
heterogeneous delay bounds. Thus, we assume all clients
require a delay bound equal to the length of a period. We
further assume that there are 6 clients in both groups.
Simulation results are shown in Figure 2. As in the case
of VoIP trafc, the modied knapsack policy achieves the
smallest total delivery debt among all the four policies.
Also, by not scheduling clients with non-positive debts,
the modied backpack policy also achieves the highest
throughput for the non-real time client.
B. Time-varying Channels
We now consider the scenario with time-varying chan-
nels, with all clients requiring delay bounds equal to
period length. We model the wireless channel by the
widely used Gilbert-Elliot model [4] [7] [19], with the
wireless channel considered as a two-state Markov chain,
with good state and bad states. A simulation study by
Bhagwat et al [2] shows that the link reliability can be
modeled as 100% when the channel is in the good state,
and 20% when the channel is in the bad state. The du-
ration that the channel stays in one state is exponentially
distributed with mean 1 10 sec for the good state, and
50 500 msec for the bad state.
While modifying the two largest debt rst policies as
suggested in Section V will yield feasibility optimality,
Fig. 3: Performance for VoIP trafc under time-varying
channels.
such modication requires solving the linear program-
ming problem and is intractable. Rather, we consider
some easier modications for the two policies. For the
largest time-based debt rst policy, we modify it so that
it treats the channel as a static one, with link reliability
equal to the time-averaged link reliability. For the largest
weighted-delivery debt rst policy, the weighted-delivery
debt for client n at time slot t is dened as
t
T
q
n
d
n
(t)
divided by the current link reliability.
For the case of VoIP trafc, we use 802.11b as the
underlying MAC and use a xed transmission rate of 11
Mb/s. We consider the same two groups of clients as in
the previous section. We assume that the mean duration
of the bad state is 500 msec for all clients, and the mean
duration of the good state is 1+0.5n sec for the n
th
client
in each subgroup. The time-average link reliability of the
n
th
client in each subgroup can be computed as
2.2+n
3+n
.
There are 19 clients in each of the subgroups.
Simulations results are shown in Figure 3. The joint
debt-channel policy incurs near zero total delivery debt,
while all the other policies have much larger total de-
livery debts. The fact that the largest time-based debt
rst policy fails to fulll the set of clients suggests that
only considering the average channel reliability, without
taking channel dynamics into account, is not satisfactory.
A somewhat surprising result is that the total delivery debt
for the largest weighted-delivery debt rst policy is even
larger than that for the random policy. This is because the
policy favors those clients with poor channels. When the
channel state is time-varying, it may make more sense
to postpone the transmissions for a client with a poor
channel until its channel condition turns better. Thus,
using weighted-delivery debt for time-varying channels is
not only inaccurate, but even harmful in some settings.
It can also be shown that the throughput for the client
with saturated non-real time trafc is the highest with
the joint debt-channel policy. By only scheduling those
real-time clients with positive delivery debts, the policy
prevents putting too much effort into any real-time client,
and thus reserves enough resources for clients with non-
realtime trafc.
For MPEG trafc, we assume there are two groups
of clients, with the same trafc patterns and delivery
ratio requirements as those in the previous section. We
use 802.11a with a xed data rate of 54 Mb/s as the
underlying MAC. The mean duration when the channel is
in the bad state is 500 msec for all clients, and the mean
duration in the good state is assumed to be 1 + 0.5n sec
for the n
th
client in each group. There are 4 clients in
both groups.
Simulation results are shown in Figure 4. As in the case
of VoIP trafc, the joint debt-channel policy incurs very
small total delivery debt while all the other policies have
9
Fig. 4: Performance for MPEG trafc under time-varying
channels
Fig. 5: Performance for VoIP trafc under heterogeneous
delay bounds
signicantly higher total delivery debts. This result sug-
gests that the simple modications of the two largest debt
rst policies do not work under time-varying channels.
Also, by only scheduling real-time clients with positive
delivery debts, the joint debt-channel policy achieves
higher throughput for the client with non-real time trafc.
C. Heterogeneous Delay Bounds
Now, we study the scenario where the channel state is
static but clients require different delay bounds. Since the
length of a period for MPEG trafc is too small, we only
simulate VoIP. There are two groups of clients. All clients
generate trafc at rate 64 kbits/sec, and thus each of them
has a packet in each period. Clients in group A require
90% delivery ratio, with delay bounds equal to the period
length. Clients in group B require 50% delivery ratio, with
delay bounds equal to two-thirds of the period length, or
22 time slots. The channel reliability for the n
th
client in
group A is (84+n)%, and that for the n
th
client in group
B is (29 + n)%.
Simulation results are shown in Figure 5. The adaptive
allocation policy has the smallest total delivery debt. This
is because the other policies, especially the two largest
debt rst policies, do not consider heterogeneous delay
bounds at all. It is not difcult to see that, to maximize
the capacity of the system, a policy should, in some
sense, work in an earliest deadline rst fashion. Without
considering heterogeneous delay bounds, the largest debt
rst policies may unwisely schedule clients with longer
delay bounds before those with shorter delay bounds, and
thus result in poor channel utilization. On the other hand,
such poor channel utilization will result in a large number
of idle time slots. Thus, the throughputs for the non-real
time trafc under these policies are higher than those for
the adaptive allocation policy.
XI. CONCLUSION
We have analytically studied the problem of scheduling
real-time trafc over wireless channels. We have extended
the model used in [10] to unreliable wireless channels
and real-time application requirements, including trafc
patterns, delay bounds, and timely-throughput bounds.
We have developed a general class of polices that are
feasibility optimal. This class can serve as a guideline
for designing computationally tractable feasibility optimal
policies. We have demonstrated the utility of the class
by deriving scheduling policies for a general case when
rate adaptation is employed and two special cases when
it is not, time-varying channels and heterogeneous delay
bounds. Simulation results show that the policies outper-
form policies described in [10]. Thus we have shown not
only that the policy class is useful in designing schedul-
ing policies, but also that neglecting some realistic and
complicated settings can result in unsatisfactory policies.
While we only present simulation results under the
IEEE 802.11 standard, the scheduling policies introduced
in this paper can be applied to a wider range to technolo-
gies including WiMax and Bluetooth. We are currently
investigating the implementation of the proposed policies
and testing them using real VoIP/MPEG trafc.
REFERENCES
[1] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whiting, and
R. Vijayakumar. Providing quality of service over a shared wireless
link. IEEE Communications Magazine, 39(2):150154, 2001.
[2] P. Bhagwat, P. Bhattacharya, A. Krishma, and S. K. Tripathi.
Using channel state dependent packet scheduling to improve TCP
throughput over wireless LANs. Wireless Networks, 3(1):91102,
1997.
[3] Y. Cao and V.O.K. Li. Scheduling algorithms in broadband wireless
networks. Proceedings of the IEEE, 89(1):7687, 2001.
[4] E. O. Elliot. Estimates of error rates for codes on burst-noise
channels. Bell Syst. Tech. J., 42:19771997, 1963.
[5] I. V. Martin F., J.J. Alins-Delgado, M. Aguilar-Igartua, and J. Mata-
Diaz. Modelling an adaptive-rate video-streaming service using
Markov-rewards models. In Proc. of QSHINE, pages 9299, 2004.
[6] H. Fattah and C. Leung. An overview of scheduling algorithms
in wireless multimedia networks. IEEE Wireless Communications,
9(5):7683, 2002.
[7] E. N. Gilbert. Capacity of a burst-noise channel. Bell Syst. Tech.
J., 39:12531265, 1960.
[8] A. Grilo, M. Macedo, and M. Nunes. A scheduling algorithm for
QoS support in IEEE802.11 networks. IEEE Wireless Communica-
tions, 10(3):3643, 2003.
[9] I-H. Hou, V. Borkar, and P.R. Kumar. A theory of QoS for wireless.
In Proc. of IEEE INFOCOM, 2009.
[10] I-H. Hou and P.R. Kumar. Admission control and scheduling
for QoS guarantees for variable-bit-rate applications on wireless
channels. In Proc. of ACM MobiHoc, pages 175184, 2009.
[11] ITU-T. Pulse Code Modulation (PCM) of voice frequencies. ITU-T
Recommendations, 1988.
[12] ITU-T. G.729 based Embedded Variable bit-rate coder: An 8-
32 kbit/s scalable wideband coder bitstream interoperable with
G.729. ITU-T Recommendations, 2006.
[13] L.J. De la Cruz and J. Mata. Performance of dynamic resources
allocation with QoS guarantees for MPEG VBR video trafc trans-
mission over ATM networks. In Proc. of GLOBECOM, pages 1483
1489, 1999.
[14] Q. Liu, X. Wang, and G.B. Giannakis. A cross-layer scheduling
algorithm with QoS support in wireless networks. IEEE Trans. on
Vehicular Technology, 55(3):839847, 2006.
[15] M. Neely. Delay analysis for max weight opportunistic scheduling
in wireless systems. In Proc. of Allerton Conf.
[16] V. Raghunathan, V. Borkar, M. Cao, and P.R. Kumar. Index policies
for real-time multicast scheduling for wireless broadcast systems.
In Proc. of IEEE INFOCOM, pages 15701578, 2008.
[17] S. Shakkottai and R. Srikant. Scheduling real-time trafc with
deadlines over a wireless channel. Wireless Networks, 8(1):1326,
2002.
[18] L. Tassiulas and A. Ephremides. Dynamic server allocation to
parallel queues with randomly varying connectivity. IEEE Trans.
on Information Theory, 39(2):89103, 1993.
[19] H.S. Wang and N. Moayeri. Finite-state Markov channel a useful
model for radio communication channels. IEEE Trans. on Vehicular
Technology, 44(1):163171, 1995.