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

4, AUGUST 2011 1125

Optimal Scheduling for Fair Resource Allocation in

Ad Hoc Networks With Elastic and Inelastic Traffic
Juan José Jaramillo, Member, IEEE, and R. Srikant, Fellow, IEEE

Abstract—This paper studies the problem of congestion control many contributions in these papers is a key modeling innovation
and scheduling in ad hoc wireless networks that have to support a whereby the network is studied in frames, where a frame is
mixture of best-effort and real-time traffic. Optimization and sto- a contiguous set of time slots of fixed duration. Packets with
chastic network theory have been successful in designing architec- deadlines are assumed to arrive at the beginning of a frame
tures for fair resource allocation to meet long-term throughput de-
mands. However, to the best of our knowledge, strict packet delay and have to be served by the end of the frame. In this paper,
deadlines were not considered in this framework previously. In this we explore this modeling paradigm further to study the design
paper, we propose a model for incorporating the quality-of-ser- of resource allocation algorithms for ad hoc networks. The
vice (QoS) requirements of packets with deadlines in the optimiza- frame-based model allows us to incorporate delay deadlines in
tion framework. The solution to the problem results in a joint con- the optimization framework for very general network models
gestion control and scheduling algorithm that fairly allocates re- and, somewhat surprisingly, allows us to design a common
sources to meet the fairness objectives of both elastic and inelastic
flows and per-packet delay requirements of inelastic flows.
framework for handling both elastic and inelastic flows.
The main contributions of the paper are as follows.
Index Terms—Ad hoc networks, best-effort traffic, congestion 1) We present an optimization framework for resource allo-
control, quality of service (QoS), real-time traffic, scheduling,
cation in a wireless network consisting of both best-effort
wireless networks.
flows and flows that generate traffic with per-packet delay
constraints. The framework allows for very general inter-
I. INTRODUCTION ference, channel, and arrival models.
2) We characterize the capacity region for wireless networks

A S WIRELESS networks become more prevalent, they

will be expected to support a wide variety of services,
including best-effort and real-time traffic. Such networks will
when there is inelastic traffic. Specifically, we show that
the capacity region not only depends on the statistics of
the channel as is the case when there is only elastic traffic,
have to serve flows that require quality-of-service (QoS) re- but it also depends on the probability distribution of the
quirements, such as minimum bandwidth and maximum delay inelastic arrivals.
constraints, while at the same time keeping the network queues 3) Using a dual decomposition approach, we derive an op-
stable for data traffic and guaranteeing throughput optimality. timal scheduling and congestion control algorithm that
For the case of wireless networks with best-effort traffic only, fairly allocates resources and ensures that a required
optimization-based algorithms that naturally map into different fraction of each inelastic flow’s packets are delivered
layers of the protocol stack have been proposed in the last few on time by appealing to connections between Lagrange
years [2]–[7]; see [8] for a survey. However, these models do multipliers, queues, and service deficits. The scheduling
not take into account strict per-packet delay bounds. algorithm seamlessly integrates inelastic and elastic traffic
Scheduling packets with strict deadlines has been studied into a unified max-weight scheduling framework, ex-
in [9]–[12], but all of these papers provide approximate solu- tending the well-known results in [16].
tions. The model that we study in this paper builds upon the 4) The convergence of the above algorithm in an appropriate
recent work in [13]–[15] on admission control and scheduling stochastic sense is proved, and it is also shown that the
for inelastic flows in collocated wireless networks, i.e., net- network is stable.
works where all links interfere with each other. Among the The rest of the paper is organized as follows. In Section II,
we present the network model, which is used in Section III
Manuscript received April 21, 2010; revised November 07, 2010; accepted to present a static optimization formulation of our problem.
November 21, 2010; approved by IEEE/ACM TRANSACTIONS ON NETWORKING
Editor A. Proutiere. Date of publication January 06, 2011; date of current Section IV presents a dual decomposition of the optimization
version August 17, 2011. This work was supported by the National Science that give us the intuition to formulate our dynamic solution
Foundation (NSF) under Grants 07-21286, 05-19691, and 03-25673; Army and prove its convergence results in Section V. Section VI
Research Office Multidisciplinary University Research Office (ARO MURI)
subcontracts; the Air Force Office of Scientific Research (AFOSR) under Grant
summarizes the differences in formulation and solution for an
FA-9550-08-1-0432; and the Defense Threat Reduction Agency (DTRA) under alternate model of the wireless channel. Section VII presents
Grant HDTRA1-08-1-0016. This work was completed during J. J. Jaramillo’s a heuristic for distributed admission control, and Section VIII
graduate studies at the University of Illinois at Urbana–Champaign. presents a simulation-based study of our algorithm. Finally,
J. J. Jaramillo was with the University of Illinois at Urbana–Champaign,
Urbana, IL 61801 USA. He is now with the Department of Electrical and Section IX presents the conclusion and future work.
Computer Engineering, Iowa State University, Ames, IA 50011 USA (e-mail:
[email protected]). II. NETWORK MODEL
R. Srikant is with the Department of Electrical and Computer Engineering and
the Coordinated Science Laboratory, University of Illinois at Urbana–Cham-
The network is represented by a directed graph ,
paign, Urbana, IL 61801 USA (e-mail: [email protected]). where is the set of nodes and is the set of directional links
Digital Object Identifier 10.1109/TNET.2010.2100083 such that, for all , if , then node
1063-6692/$26.00 © 2011 IEEE
1126 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 4, AUGUST 2011

can transmit to node . The links are numbered 1 through ,

and by abusing notation, we sometimes use to mean
.
Traffic is assumed to be a mixture of elastic and inelastic
flows, where an inelastic flow is one that has maximum per-
packet delay requirements. In contrast, elastic flows do not have
such requirements.
Time is divided in slots, where a set of consecutive time Fig. 1. Example of a feasible schedule when a = (1; 0; 1) and c = (1; 1; 1).
slots makes a frame. We assume that packet arrivals only occur
at the beginning of a frame, and every inelastic packet has a
deadline of time slots. If a packet misses its deadline, it is dis- sake of simplicity, we will use a channel and interference model
carded, and it is required that the loss probability at link that allow us to present the main ideas behind our algorithm
due to deadline expiry must be no more than . For elastic without the need for more complex notation.
traffic, we associate a utility function , which is a func- In the rest of this paper, we will consider the known channel
tion of the mean elastic arrival rate per frame . We assume case first, and then in Section VI we will highlight the differ-
that is a concave function. ences in the analysis for the unknown channel case.
For a given frame, we denote by the vector the
number of inelastic packet arrivals at every link, where is III. PROBLEM FORMULATION
a random variable with mean and variance . We further
assume that arrivals are independent between different frames We first formulate the problem as a static optimization
and that and . The last two problem. Using decomposition theory, we will then obtain a
assumptions are used to guarantee that the Markov chain we dynamic solution to this problem and prove its stability using
define later is both irreducible and aperiodic, although these can stochastic Lyapunov techniques.
be replaced by other similar assumptions. Similarly, we define A feasible schedule is such that ,
to be the number of elastic packet arrivals at every respectively denote the number of inelastic and elastic packets
link in a given frame. that can be scheduled for transmission at link and time
The channel state is assumed to be constant in a given frame, slot . Thus, means that link
independent between different frames and independent of ar- is scheduled to transmit in time slot of the frame. Furthermore,
rivals. The vector denotes the number of packets for any , if and , then links
link can successfully transmit on a time slot in a given frame. and can be scheduled to simultaneously transmit without in-
Depending on the wireless technology used, we can have terfering with each other. Assuming the inelastic arrivals and
some channel feedback before or after a transmission occurs. If the channel state are given by and , respectively, we have
channel estimation is performed before transmitting, we can de- the following constraints:
termine the optimal rate at which we can successfully transmit.
Alternatively, feedback from the receiver after the transmission for all (1)
can be used to detect if a transmission is successful or not. In this
paper, we try to capture both scenarios in the following cases. for all and (2)
1) Known channel state: It is assumed that is a nonnegative
random variable with mean and variance , and we get Since an inelastic packet that misses its deadline is discarded,
to know the channel state at the beginning of the frame. (1) states that the number of successful inelastic transmissions
2) Unknown channel state: It is assumed that is a Bernoulli within a frame is bounded by the number of packet arrivals. In
random variable with mean , and we only get to know the other words, there is no concept of queueing inelastic packets
channel state at the end of the frame. after deadline expiration. Similarly, (2) states that the number of
In the known channel state case where we do channel estima- successful transmissions in any slot cannot exceed the state of
tion to determine the optimal transmission rate, we can poten- the channel. We denote by the set of all feasible sched-
tially send more than one packet in a time slot at higher rates. ules when the arrival state is and the channel state is . Thus,
This is captured by the fact that we make no assumptions on the captures any interference constraints we have on our
values can take since it will be determined by the particular network and satisfies (1) and (2).
wireless technology used. In the case of unknown channel state, As an example of a feasible schedule, consider Fig. 1. We
we assume that we only get the binary feedback of acknowledg- consider a network with three links such that only one link
ments, which is reflected in the Bernoulli assumption on . In can be scheduled at any given time and assume that a frame is
this case and without any loss of generality, we assume only one composed of five time slots. Furthermore, we assume that the
packet can be transmitted per time slot per link. channel is always ON and only one packet can be transmitted
The reader is referred to [17] for an alternative method to at any given time, so , and that the inelastic ar-
define more general channel and interference models, where rivals are given by the vector . Since only one link
the channel is modeled by a reliability function that captures can be scheduled at any given time, we know that a feasible
the probability of successful transmission as a function of the schedule must satisfy for all time slots
channel state and the physical-layer transmissions parameters. in a frame. Thus, it is easy to see that our schedule meets the
It must be noted, though, that having more general models does interference constraints and (2). In the case of link 2, we have
not change the nature of the solution to our problem, so for the that , meeting constraint (1). On
JARAMILLO AND SRIKANT: OPTIMAL SCHEDULING FOR FAIR RESOURCE ALLOCATION IN AD HOC NETWORKS 1127

the other hand, since elastic packets can be queued during mul- We will focus on maximizing the following objective for
tiple frames, we can schedule their transmission independent of some given vector :
the current queue length. Therefore is a valid
schedule. A similar analysis can be done to verify that (1) is also
satisfied for links 1 and 3. (3)
At the beginning of any frame, we must choose a feasible
schedule to serve all links and decide how many elastic packets subject to
are allowed to be injected in the network. Therefore, our goal
is to find a function that is the probability of using for all
schedule when the inelastic arrivals are given by for all
and the channel state is , subject to the constraint that the loss
probability at link due to deadline expiry cannot exceed The vector can be used to allocate additional bandwidth
. For elastic traffic, we want to select the vector such that we fairly to inelastic flows beyond what is required to meet their
maximize the network utility while keeping the queues stable. QoS needs. Other uses for will be explored in the simula-
To properly formulate the problem, let us first define tions section. We will assume that the inelastic arrivals and loss
to be the expected number of inelastic packets served if the probability requirements are feasible, and thus the optimization
number of packet arrivals is given by and the channel state problem has a solution .
is . Similarly, denotes the expected number of elastic
packets that can be served. Therefore, we have the following IV. SOLUTION USING DUAL DECOMPOSITION
constraints: Using the definition of the dual function [18], we have that

subject to
for all
The expected service for mixed traffic at link is then given
by Slater’s condition [19] states that since the objective is con-
cave and the constraints are affine functions, the duality gap is
zero, and therefore , where

We are interested in finding , but not the value

and due to QoS requirements and capacity constraints, we re- , so if we rewrite the objective in the dual function as
quire that

Furthermore, to avoid trivialities, we will assume that we notice that the problem can be decomposed into the fol-
for all . lowing subproblems:
If we define the capacity region for fixed arrival and channel
states and as follows:
there exists
and
and
(4)

it is easy to see that Furthermore, since we are interested in solving the problem
, where is the convex hull of . for nonnegative values of and , it must be the case that
Similarly, if the overall capacity of the network is defined as and are as large as the constraints allow, and since the
upper bounds for and are expressed as a
convex combination and the objective function in (4) is linear,
there exists the problem can be decomposed into the following subproblems
for all and for fixed and :
for all

we have that .
1128 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 4, AUGUST 2011

This suggests the following iterative algorithm to find the so- V. DYNAMIC ALGORITHM AND ITS CONVERGENCE ANALYSIS
lution to our optimization problem, where is the step index
and is a large enough parameter: A. Scheduler and Congestion Controller
To implement the algorithm online, we propose the following
congestion control algorithm in frame , where the queue length
at link is given by :

(5)

We need to convert this elastic arrival rate, which in general

is a nonnegative real number, into a nonnegative integer indi-
cating the number of elastic packets allowed to enter the net-
work in a given frame. This conversion can be made in many dif-
ferent ways. We assume the elastic arrivals at link , , are a
random variable with mean and variance upper-bounded
by and are such that and
We update the Lagrange multipliers , at every step for all and all . The last two assumptions are
according to the following equations: used to guarantee the Markov chain we define below is both ir-
reducible and aperiodic, although these can be replaced by other
similar assumptions.
Letting the number of inelastic arrivals be denoted by
and and the channel state by , we propose the following sched-
uling algorithm:

where is a fixed step-size parameter, and for any ,

.
Making the change of variables and (6)
, we have that our iterative algorithm can be rewritten as
The vectors and are updated from frame to frame
as follows:

where

with update equations

and is a binomial random variable with parameters
and . The quantity can be generated by the network
as follows: Upon each inelastic packet arrival, toss a coin with
probability of heads equal to , and if the outcome is heads,
add a 1 to the deficit counter.
It should be noted that due to the change of variables, In our notation, we make explicit the fact that for fixed and
can be interpreted as a queue that has arrivals and , the optimal scheduler (6) is a function of , , ,
departures at step ; can have a similar queue in- and . We interpret as a virtual queue that counts the
terpretation. The dual decomposition approach only provides an deficit in service for link to achieve a loss probability due to
intuition behind the solution, but the real network has stochastic deadline expiry less than or equal to . This deficit queue was
and dynamic arrivals and channel state conditions. In Section V, first used in the inelastic traffic context in [13] for the case of
we present the complete solution that takes into account these collocated networks; the connection to the dual decomposition
dynamics, and we also establish its convergence properties. approach now provides a Lagrange multiplier interpretation to
JARAMILLO AND SRIKANT: OPTIMAL SCHEDULING FOR FAIR RESOURCE ALLOCATION IN AD HOC NETWORKS 1129

it and allows the extension to general ad hoc networks. Note that Lemma 1 also implies that the scheduling and congestion
is just the queue size for elastic packets at link . control algorithm fulfills all links’ inelastic requirements.
Corollary 2: If there exists a point for
B. Convergence Results some such that (7) and (8) hold true, then the online
algorithm fulfills all the inelastic constraints. That is
For readability, we present the main results in this section,
but the proofs are deferred to the appendixes. We start by noting
that defines an irreducible and aperiodic Markov
chain. We will first bound the expected drift of for
a suitable Lyapunov function.
Lemma 1: Consider the Lyapunov function
for all .
. If there exists a point for
The above corollary simply states that the arrival rate into the
some such that
deficit counter is less than or equal to the departure rate. This
result is an obvious consequence of the stability of the deficit
for all (7)
counters, and so a formal proof is not provided here.
To prove that our dynamic algorithm achieves the optimal
and solution to the static problem (3) in some average sense and
fulfills all links’ requirements, we first state a weaker result than
for all (8) Lemma 1.
Lemma 2: Consider the Lyapunov function
then . Then

for some nonnegative constants , , , any ,

where is the solution to (3), is the solution to (5),
for some positive constants , , , any , where and is obtained from the solution to (6).
is the solution to (5) and is obtained from The proof follows the same lines as in Lemma 1 and is thus
the solution to (6). omitted.
It is important to note that since the last two terms in the right- Now, we are ready to prove that our online algorithm is within
hand side of the inequality can be upper-bounded, Lemma 1 of the optimal value.
implies that is positive recurrent since the expected Theorem 1: For any and assuming that is a con-
drift is negative but for a finite set of values of . As cave function, we have that
a direct consequence of this fact, we note that the total service
deficit and queue length have a bound.
Corollary 1: If there exists a point
for some such that (7) and (8) hold true, then the
total expected service deficit and network queue length is
upper-bounded by

for some , where is the solution to (3),

is the solution to (5), and is obtained
from the solution to (6).
It is interesting to note that the assumption that is a
where
concave function is only used to prove that our algorithm is close
to the optimal value, but it is not necessary to prove any of the
other results in this section.
In conclusion, there is a tradeoff in choosing the parameter :
and Smaller values will achieve a solution closer to the optimal, but
at the same time the deficit in service at the links and the ag-
gregate queue length increase. The statement and the proofs of
Lemma 1 and Theorem 1 follow techniques similar to those
in [4]. The results can also be derived, in a slightly different
form, using the techniques in [5]. A closely related result can be
obtained using the methods in [2].
1130 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 4, AUGUST 2011

VI. UNKNOWN CHANNEL STATE To avoid trivialities, we will assume that for all
.
The analysis for the unknown channel case is similar to the
If we define the capacity region for fixed arrival and channel
one we presented for the known channel case, so in this section
states and as follows:
we will only highlight the differences.
A feasible schedule is such that , there exists
respectively denote the number of inelastic and elastic packets
that can be scheduled for transmission at link and and
time without violating any interference
constraints. Assuming the inelastic arrivals are given by and
it is easy to check that
since we can only schedule at most one packet per link at every
, where is the convex hull of .
time slot, we have the following constraints:
Similarly, the overall capacity of the network can be defined
as
for all (9)

for all and (10) there exists

for all and
We denote by the set of all feasible schedules for fixed for all
arrivals, capturing any interference constraints we have on our
network and satisfying (9) and (10). we have that .
Our goal now is to find a function , which is the prob- With the definitions and constraints stated above, we can for-
ability of using schedule when the inelastic arrivals mulate the optimization problem in a similar way as in (3) for
are given by , subject to the constraint that the loss probability some vector
at link due to deadline expiry cannot exceed . For elastic
traffic, we still want to select the vector such that we maxi-
mize the total utility while keeping the queues stable.
For a given distribution , we have that is the
expected number of attempted inelastic transmissions if arrivals subject to
are given by . Similarly, denotes the expected number
of times link is scheduled to serve elastic packets in a given for all
frame. As before, we have the following constraints: for all

Using the dual decomposition approach outlined in

Section IV, we find that the only difference with the known
channel state case is the scheduling algorithm. Assuming
inelastic arrivals are given by , the scheduling algorithm
is given by

When the (unknown) channel state is , we have that

is the expected number of successful inelastic transmissions per
frame at link for fixed arrivals, while is the expected
service to link for inelastic arrivals. Thus, the expected service
for mixed traffic at link is given by
The main difference in the scheduling algorithm compared to
the known channel state case is that the network now uses the
expected channel state in making scheduling decisions. Thus,
the network needs to know or estimate as in [13].
Similar results can be proven for this algorithm using the tech-
niques developed in Section V-B, whereby one can show that
the algorithm meets all the inelastic QoS constraints, the total
Simplifying both expressions, we get expected service deficits and the queue lengths have a
bound, and the mean value of the objective is within of the
optimal value.

VII. DISTRIBUTED ADMISSION CONTROL

In this section, we present a heuristic for distributed admis-
Due to service requirements and capacity constraints, we sion control based on the fact that the deficit counters are
need that a good indicator of stability in the network.
So far, we have proven that if the inelastic arrivals and
loss probability requirements are feasible, then our solution
JARAMILLO AND SRIKANT: OPTIMAL SCHEDULING FOR FAIR RESOURCE ALLOCATION IN AD HOC NETWORKS 1131

Fig. 2. Interference graph used in the simulations.

stabilizes the network and achieves an optimal allocation of re-

sources without explicitly needing to characterize the capacity
region. In the case that the inelastic arrivals and loss probability
requirements are not necessarily feasible, we need to develop
an admission control mechanism that guarantees that the set Fig. 3. Deficit size and queue length when w = 0.
of admitted flows can be supported by the network. Since the
congestion controller (5) guarantees that the rate of elastic
packets injected in the network are always within the capacity
region, we only need to decide on the admission of inelastic
flows. A possible solution would be to characterize the capacity
region analytically and use an admission controller that exploits
this characterization to admit or reject new flows. However,
a simple characterization of the capacity region appears to be
difficult for ad hoc networks, and furthermore, any controller
that uses such a characterization will be centralized in any case
and is thus not very useful.
In Corollary 1, we proved that if the requirements can be sup-
ported by the network using the service vector , then the
expected service deficit is upper-bounded. Thus, if we see that
the deficit counter of an inelastic flow does not stabilize over
time, it means that such flow cannot be supported by the net-
work. This observation can be exploited to perform admission Fig. 4. Deficit size and queue length when w = 3.
control as follows. When a new flow arrives at a link, it can ask
its neighbors for the values of their deficit queues. Admit the
new flow only if the neighbors’ deficit counters are all below a every frame. For large networks, this can become a burden due
threshold delta. The proper choice of such a delta and whether to the large search space. Thus, we explore a greedy heuristic
one can prove desirable properties of such an admission con- and check how close it is to the optimal solution. At any given
troller are topics for further research. time slot, the greedy scheduler orders all links according to their
weights. The greedy scheduler adds one of the links with the
largest weight to the schedule, then removes all links that inter-
VIII. SIMULATIONS
fere with this link from the graph, then schedules a link with the
The purpose of this simulation study is to understand how largest weight among the remaining links, and so on. This pro-
the parameter and the link weights impact the performance cedure continues until no more links can be scheduled.
of the algorithm, how a greedy heuristic can be used to imple- In Figs. 3–5, we plot the expected values of the deficit coun-
ment the optimal scheduler, and how channel information helps ters and queues per link for various values of and compare
improve performance. We simulate a 10-link network with an their evolution for both the scheduler with optimal decisions and
interference graph given by Fig. 2, where each node represents the greedy scheduler.
a link and each edge means that the two adjacent links cannot We see that as increases, the deficit counters become small.
be scheduled simultaneously. For example, if link 1 is sched- The upper bound in Corollary 1 only suggests that the sum of the
uled, then links 2, 4, and 7 cannot be activated. The required deficit counters and queues is . Thus, it is interesting to
loss probability due to deadline expiry of inelastic packets is set note that by changing , one can nearly eliminate the backlog
to 0.1, the link arrivals are assumed to have a Bernoulli distri- in deficit for inelastic traffic while maintaining the same order
bution with mean 0.6 packets/frame, and there are three time of queue sizes. The reason for this can be understood by ex-
slots per frame. The channel for every link is assumed to have amining the scheduling algorithm (6). Note that the algorithm
a Bernoulli distribution with mean 0.96, and for the first part of gives priority to elastic traffic if queues are larger than counters.
the simulations, we assume we get to know the channel state at When is small compared to , the effect of is negligible in
the beginning of the frame. We set for all the scheduling algorithm. On the other hand, when is ,
links. The simulation time was 10 frames. is , which is comparable to the queue lengths, and
As can be noted from (6), the max-weight scheduler requires hence the deficit does not have to be large to provide service to
that we do an exhaustive search to find the optimal schedule at inelastic traffic under algorithm (6).
1132 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 4, AUGUST 2011

Fig. 5. Deficit size and queue length when w = 6. Fig. 7. Average service when w = 3.

Fig. 6. Average service when w = 0. Fig. 8. Average service when w = 6.

It must be noted that small deficit counters mean that there is

a small backlog in providing acceptable service to inelastic ar-
rivals. For the case of real-time traffic, this is a desirable property
since we do not want to have large variations in the service pro-
vided that could affect the perceived quality. Thus, even if fair
allocation of bandwidth beyond the minimum is not required for
inelastic flows, choosing an order of magnitude larger than
is desirable to maintain small deficits.
As can be noted, the greedy scheduler seems to give lower
deficit values than the optimal scheduler for larger values of .
We believe that the reason is that weights given to inelastic flows
increase with increasing , and therefore the greedy scheduler
picks them first. However, our optimality goal is given by (3),
which is determined by the rates received by the various flows.
The rates achieved by the two schedulers are quite close in the Fig. 9. Dropping probability when w = 0.
simulations, as seen in Figs. 6–8, while keeping the dropping
probabilities below the requirement, as shown in Figs. 9–11.
Finally, we study the loss of performance when we get IX. CONCLUSION AND FUTURE WORK
to know the channel state at the end of the frame compared In this paper, we have presented an optimization framework
to knowing the channel at the beginning of the frame. In for the problem of congestion control and scheduling of elastic
Figs. 12–14, we see that the rates achieved for inelastic traffic and inelastic traffic in ad hoc wireless networks. The model was
are identical, while elastic traffic has a performance loss no developed for general interference graphs, general arrivals, and
greater than 10% in any case. The reason for this comes from time-varying channels. Using a dual-function approach, we pre-
the fact that inelastic traffic has to meet throughput constraints, sented a decomposition of the problem into an online algorithm
while elastic traffic uses the excess capacity. Thus, any loss in that is able to make optimal decisions while keeping the network
capacity is mostly absorbed by elastic traffic. stable and fulfilling the inelastic flow’s QoS constraints. A key
JARAMILLO AND SRIKANT: OPTIMAL SCHEDULING FOR FAIR RESOURCE ALLOCATION IN AD HOC NETWORKS 1133

Fig. 10. Dropping probability when w = 3. Fig. 13. Average service when w = 3.

Fig. 11. Dropping probability when w = 6. Fig. 14. Average service when w = 6.

The case when the channel state varies from time slot to time
slot is more challenging to solve since it involves a dynamic
programming solution. Similarly, our traffic model for inelastic
packets assumes that packets arrive at the beginning of the frame
and all have the same delay. Possible extensions include hetero-
geneous delays and arrivals at any time slot. These topics are
left for future work.

APPENDIX A
PROOF OF LEMMA 1
To prove Lemma 1, we start by first proving two auxiliary
lemmas and then stating a fact.
Lemma 3: Given that at frame we have the event
and and if there exists a point
Fig. 12. Average service when w = 0. for some such that (7) and (8) hold true, then

result is that, through the use of deficit counters, one can treat the
scheduling problem for elastic and inelastic flows in a common
framework. It is also interesting to note that the deficit coun-
ters introduced in [13]–[15] have the interpretation of Lagrange
multipliers. Simulations corroborate our results and show the
dependency of the performance of the algorithm on the auxil-
iary parameter and its role into assigning resources to both
elastic and inelastic traffic.
It must be noted that in the unknown channel state case, we as-
sume that the channel remains constant during the entire frame.
1134 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 4, AUGUST 2011

for some nonnegative constant and where for some constant , where and are the solutions
is given by the solution to (6). to (3) and (5), respectively.
Proof: Proof:

(11)
(13)

(14)
(12)

where (13) follows from the definition of

where (11) follows from the definition of ,
(12) follows from (7), and

and (14) follows from the fact that is the optimal point
of (5).
Fact 1: The optimization in (6) can be performed over
Lemma 4: Given that at frame we have the event
, the convex hull of ; that is
and and if there exists a point
for some such that (7) and (8) hold true, then
JARAMILLO AND SRIKANT: OPTIMAL SCHEDULING FOR FAIR RESOURCE ALLOCATION IN AD HOC NETWORKS 1135

The reason for this comes from the fact that the objective where ; (15) follows from the definition of
function is linear, and therefore there must be an optimal point , , , and Fact 1. Thus, we have the
. following:
Proof of Lemma 1: From the definition of , we
know that means that there exist
for all , and
,
for all . For the rest of the proof, we define ,
to be such set of values associated to .
From Lemmas 3 and 4, we have

where

and

APPENDIX B
PROOF OF THEOREM 1
From Lemma 2, we know that

since . Taking expectations

Adding the terms for and dividing by , we

get

(15) (16)
1136 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 4, AUGUST 2011

where (16) follows from the fact that the Lyapunov function [9] S. Shakkottai and R. Srikant, “Scheduling real-time traffic with dead-
is nonnegative. Using Jensen’s inequality [19], we get lines over a wireless channel,” Wireless Netw., vol. 8, no. 1, pp. 13–26,
Jan. 2002.
[10] V. Raghunathan, V. Borkar, M. Cao, and P. R. Kumar, “Index policies
for real-time multicast scheduling for wireless broadcast systems,”
in Proc. IEEE INFOCOM, Phoenix, AZ, Apr. 13–18, 2008, pp.
1570–1578.
[11] A. Dua and N. Bambos, “Downlink wireless packet scheduling
with deadlines,” IEEE Trans. Mobile Comput., vol. 6, no. 12, pp.
1410–1425, Dec. 2007.
[12] Q. Liu, X. Wang, and G. B. Giannakis, “A cross-layer scheduling al-
gorithm with QoS support in wireless networks,” IEEE Trans. Veh.
Technol., vol. 55, no. 3, pp. 839–847, May 2006.
[13] I.-H. Hou, V. Borkar, and P. R. Kumar, “A theory of QoS for wireless,”
in Proc. IEEE INFOCOM, Rio de Janeiro, Brazil, Apr. 19–25, 2009,
pp. 486–494.
[14] I.-H. Hou and P. R. Kumar, “Admission control and scheduling for QoS
guarantees for variable-bit-rate applications on wireless channels,” in
Proc. 10th ACM MobiHoc, New Orleans, LA, May 18–21, 2009, pp.
175–184.
[15] I.-H. Hou and P. R. Kumar, “Scheduling heterogeneous real-time traffic
over fading wireless channels,” in Proc. IEEE INFOCOM, San Diego,
Assuming , we get the following limit CA, Mar. 15–19, 2010, pp. 1–9.
expression: [16] L. Tassiulas and A. Ephremides, “Stability properties of constrained
queueing systems and scheduling policies for maximum throughput in
multihop radio networks,” IEEE Trans. Autom. Control, vol. 37, no. 12,
pp. 1936–1948, Dec. 1992.
[17] M. J. Neely, “Delay-based network utility maximization,” in Proc.
IEEE INFOCOM, San Diego, CA, Mar. 15–19, 2010, pp. 1–9.
[18] D. G. Luenberger, Linear and Nonlinear Programming, 2nd ed. Nor-
well, MA: Kluwer, 2003.
[19] S. Boyd and L. Vandenberghe, Convex Optimization, 1st ed. New
York: Cambridge Univ. Press, 2004.
where .
Juan José Jaramillo (S’06–M’11) received the B.S.
degree (summa cum laude) from Universidad Ponti-
ACKNOWLEDGMENT ficia Bolivariana, Medellin, Colombia, in 1998, and
This paper is a revised version of an earlier paper that ap- the M.S. and Ph.D. degrees from the University of
Illinois at Urbana–Champaign in 2005 and 2010, re-
peared in [1]. The main differences between this version and spectively, all in electrical engineering.
the previous one are that the proofs have been rewritten to make From 1999 to 2003, he was with Empresas
the results more general and simulations have been extensively Publicas de Medellin, Medellin, Colombia. He is
expanded. currently with Iowa State University, Ames, where
he is a Post-Doctoral Research Associate with the
Department of Electrical and Computer Engineering.
REFERENCES He is the recipient of a Fulbright fellowship. His research interests include
[1] J. J. Jaramillo and R. Srikant, “Optimal scheduling for fair resource communication networks and game theory.
allocation in ad hoc networks with elastic and inelastic traffic,” in Proc.
IEEE INFOCOM, San Diego, CA, Mar. 15–19, 2010, pp. 1–9.
[2] A. Eryilmaz and R. Srikant, “Fair resource allocation in wireless net-
works using queue-length-based scheduling and congestion control,” R. Srikant (S’90–M’91–SM’01–F’06) received
in Proc. IEEE INFOCOM, Miami, FL, Mar. 13–17, 2005, vol. 3, pp. the B.Tech. degree from the Indian Institute of
1794–1803. Technology, Madras, India, in 1985, and the M.S.
[3] X. Lin and N. B. Shroff, “Joint rate control and scheduling in multihop and Ph.D. degrees from the University of Illinois at
wireless networks,” in Proc. 43rd IEEE CDC, Atlantis, Bahamas, Dec. Urbana–Champaign in 1988 and 1991, respectively,
14–17, 2004, vol. 2, pp. 1484–1489. all in electrical engineering.
[4] M. J. Neely, E. Modiano, and C.-P. Li, “Fairness and optimal He was a Member of Technical Staff with AT&T
stochastic control for heterogeneous networks,” in Proc. IEEE IN- Bell Laboratories, Holmdel, NJ, from 1991 to 1995.
FOCOM, Miami, FL, Mar. 13–17, 2005, vol. 3, pp. 1723–1734. He is currently with the University of Illinois at Ur-
[5] A. Stolyar, “Maximizing queueing network utility subject to stability: bana-Champaign, where he is the Fredric G. and Eliz-
Greedy primal-dual algorithm,” Queue. Syst., vol. 50, no. 4, pp. abeth H. Nearing Endowed Professor with the De-
401–457, Aug. 2005. partment of Electrical and Computer Engineering and a Research Professor in
[6] A. Eryilmaz and R. Srikant, “Joint congestion control, routing and mac the Coordinated Science Laboratory. His research interests include communi-
for stability and fairness in wireless networks,” IEEE J. Sel. Areas cation networks, stochastic processes, queueing theory, information theory, and
Commun., vol. 24, no. 8, pp. 1514–1524, Aug. 2006. game theory.
[7] L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Cross-layer con- Prof. Srikant was an Associate Editor of Automatica and the IEEE
gestion control, routing and scheduling design in ad hoc wireless TRANSACTIONS ON AUTOMATIC CONTROL and is currently an Associate Editor
networks,” in Proc. IEEE INFOCOM, Barcelona, Spain, Apr. 23–29, of the IEEE/ACM TRANSACTIONS ON NETWORKING. He has also served on the
2006, pp. 1–13. Editorial Boards of special issues of the IEEE JOURNAL ON SELECTED AREAS
[8] X. Lin, N. B. Shroff, and R. Srikant, “A tutorial on cross-layer opti- IN COMMUNICATIONS and IEEE TRANSACTIONS ON INFORMATION THEORY.
mization in wireless networks,” IEEE J. Sel. Areas Commun., vol. 24, He was the Chair of the 2002 IEEE Computer Communications Workshop in
no. 8, pp. 1452–1463, Aug. 2006. Santa Fe, NM, and was a Program Co-Chair of IEEE INFOCOM 2007.