IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 3, MARCH 2011
1161
Availability-Aware Multimedia Scheduling in
Heterogeneous Wireless Networks
Liang Zhou, Haohong Wang, Shiguo Lian, Yan Zhang, Athanasios V. Vasilakos, and Weiping Jing
Abstract—Scheduling multimedia applications over heterogeneous wireless networks is a challenging issue due to qualityof-service (QoS) requirements, different resource requests, and
dynamically available network resources. Resource availability is
an important constraint for the adaptive usage of limited network
resources. However, it has not been well studied in the literature. In
this paper, we propose a novel distributed availability-aware adaptive rate-allocation scheme for multimedia applications. A general
distortion model is first proposed, which is highly dependent on the
application’s rate-distortion characteristics and the availability
model. Then, a near-optimal rate-allocation approach is presented
that jointly considers QoS, reliability, and availability. Numerical
results indicate that the proposed scheme is able to achieve higher
QoS under various environments compared with other reference
approaches.
Index Terms—Availability aware, heterogeneous wireless
networks, multimedia applications, quality-of-service (QoS),
scheduling.
I. I NTRODUCTION
R
ECENTLY, we have witnessed the increasing efforts toward standardization of architectures for convergence of
heterogeneous access networks. In addition, the integration of
heterogeneous networks is becoming an integral part of the
fourth-generation network design [1]. Supporting multimedia
applications over heterogeneous networks is an inherent requirement as well as a challenging issue in the communications and multimedia research communities. For example, the
internet protocol Multimedia Subsystems (IMS) platform [2]
Manuscript received October 19, 2009; revised March 23, 2010 and
September 10, 2010; accepted December 6, 2010. Date of publication
January 10, 2011; date of current version March 21, 2011. This work was
supported in part by the Alexander von Humboldt Foundation, by the Science
Technology Project of the Ministry of Transport of the People’s Republic of
China under Grant 2009-353-332-290, and by the AURORA project of the
Research Council of Norway under Grant 205048/V11. The review of this paper
was coordinated by Prof. H. Hassanein.
L. Zhou is with the Jiangsu Province Key Laboratory of ASIC Design,
Nantong University, Nantong 226019, China, and also with the Technische Universität München, 80333 München, Germany (e-mail: liang.zhou@ieee.org).
H. Wang is with Cisco Systems, San Jose, CA 95134 USA (e-mail:
haohong@ieee.org).
S. Lian is with France Telecom R&D (Orange Laboratories), Beijing 100080,
China (e-mail: sglian@gmail.com).
Y. Zhang is with the Simula Research Laboratory, 1325 Lysaker, Norway
(e-mail: yanzhang@simula.no).
A. V. Vasilakos is with the Department of Computer and Telecommunications Engineering, University of Western Macedonia, 50100 Kozani, Greece
(e-mail: vasilako@ath.forthnet.gr).
W. Jing (Corresponding author) is with the Jiangsu Province Key Laboratory of ASIC Design, Nantong University, Nantong 226019, China (e-mail:
jingwp@yahoo.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2011.2104420
Fig. 1. Example of multimedia streaming architecture in heterogeneous wireless networks.
has defined an overlay architecture for providing multimedia
services on top of heterogeneous wireless networks.
Fig. 1 shows the scenario where multiple users transmit
multimedia applications with different quality-of-service (QoS)
over existing heterogeneous wireless networks [3]. In this architecture, there are various kinds of application data streaming
from different users. Hence, a rate-allocation policy plays an
important role in efficiently using network resources. Compared with traditional networks, the rate-allocation problem
over heterogeneous networks is more complicated by the heterogeneities of the application contents and the network dynamics. The performance of a heterogeneous system may be
degraded if one of the networks is out of order in case of random
or preventive maintenance. Meanwhile, many media applications require transmission platform with high availability. As
such, a rate-allocation strategy in heterogeneous systems should
consider availability information to deal with unexpected failures [4]. In the literature, there are several scheduling schemes
for multimedia applications in wireless networks (e.g., [5], [7]–
[11], [20], and [21]). In these studies, availability information
has not been sufficiently taken into account when making resource allocation decisions, and thus, they may not be appropriate or efficient for availability critical multimedia applications
[12]. Hence, availability-aware rate-allocation schemes should
be developed to satisfy the need for high quality of availability
demanded by multimedia applications. Motivated by this, in
this paper, we design and evaluate an availability-aware rateallocation scheme by jointly considering media applications,
heterogeneous networks reliability, and service availability.
The issue of rate allocation among multiple traffic flows
over shared network resources has been receiving increasing
0018-9545/$26.00 © 2011 IEEE
1162
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 3, MARCH 2011
attention recently. Internet applications typically use the Transmission Control Protocol (TCP) congestion control mechanism
to regulate the outgoing rate [13], [17]. For media-streaming
applications over the User Datagram Protocol, TCP-Friendly
Rate Control is a popular choice [14], and several modifications
have been proposed to improve its media friendliness [15].
The problem of efficient utilization of multiple networks via
a suitable allocation of traffic flows has also been explored
from different perspectives. A queuing-based dynamic channel
selection for heterogeneous multimedia applications has been
presented in [7]. A game-theoretic framework to allocate bandwidth for elastic services in networks with fixed capacities has
been addressed in [16]. In addition, the rate adaptation of multimedia streams has been studied in the context of heterogeneous
networks in [18], where the authors propose an architecture to
allow online measurement of network characteristics and video
rate adaptation via transcoding. In [19], the media-aware rate
allocation has been achieved, taking into account the impact
of both packet-loss rates and available bandwidth over each
link, on the end-to-end video quality of a single stream. In
[8], [9], and [11], the rate-allocation problem is considered
in a heterogeneous environment by taking into account the
rate-distortion characteristics. To the best of our knowledge,
our paper is the first one to provide dependable and different
media applications that meet the users’ requirements in terms
of QoS, reliability, and availability. The major contributions in
this paper include the following:
1) the design and evaluation of a distributed scheduling scheme for multiple real-time media applications
with availability constraints running on heterogeneous
networks;
2) the proposition of a general media distortion model that
can be used to quantitatively measure the performance of
scheduling in terms of the distortion;
3) the investigation of the optimal or close-to-optimal rateallocation scheme under the resource constraints.
The rest of this paper is organized as follows: We present
our system model and problem formulation in Section II. In
Section III, we propose an availability-aware adaptive scheduling scheme for multiple-application streaming sessions sharing multiple heterogeneous networks. Then, we provide some
selected simulation results for the proposed rate-allocation
scheme in Section IV. Section V concludes this paper and
points to future work.
II. S YSTEM M ODEL
Here, we present the service model, heterogeneous wireless
networks, and availability model, which are built to represent a
system distortion framework.
A. Multiapplication Service
We suppose that multiple users S = {1, . . . , s, . . . , S} simultaneously access one of two different applications: realtime video streaming (A1 ) and audio conversation (A2 ). Let
user s ∈ S access one of the available applications Ai (i =
1, 2). The two applications are stored in a server, and the
server determines the allocated rate to user s that has chosen application Ai . We assume that the server can scalably
adapt the transmission process to the channel conditions for
user s. The server can choose the appropriate transmission
parameters from a predefined set of available parameters RAi
for the application Ai . There are NA1 encoded video layers and NA2 audio transcoders available at the server. Each
video layer l (l ≤ NA1 ) is characterized by its average encoding rate ρl , and each transcoder v (v ≤ NA2 ) is characterized by its encoding rate ρv . We define RA1 = {ρl :
1 ≤ l ≤ NA1 } and RA2 = {ρv : 1 ≤ v ≤ NA2 } as the sets of
available parameters for the video and audio applications,
respectively.
In practice, all the video and audio sources should be compressed for transmission and storage, and the compression may
lead to some information loss. The distortion caused by source
compression Dcomp can be approximated by [20]
Dcomp =
θ
+ D0
R − R0
(1)
where R is the rate of the video/audio stream, which is equal
to RAi (i = 1, 2); θ, R0 , and D0 are the parameters of the distortion model that depend on the encoded video/audio sequence
and on the encoding structure. Using nonlinear regression techniques, these parameters can be estimated from empirical ratedistortion curves obtained by encoding a sequence at different
rates [21].
B. Heterogeneous Wireless Networks
In this section, we present the distortion model due to packet
loss in heterogeneous wireless networks. Similar to Dcomp , the
distortion caused by packet loss can be modeled by a linear
model related to the packet-loss rate Ploss , i.e.,
Dloss = κPloss
(2)
where κ depends on the parameters related to the compressed
sequence, e.g., the proportion of intracoded macroblocks and
the effectiveness of error concealment at the decoder [20]. The
packet-loss rate Ploss represents the combined rate of random
losses and expired arrivals of packets. In a bandwidth-limited
wireless network, this combined loss rate can be modeled by
an M/G/1 queuing model. In this case, the packet delay over a
single link follows an exponential distribution [21]. Since the
packet end-to-end delay in a wireless network is dominated by
the queuing delay at the bottleneck link, the empirical delay
distribution for realistic traffic can still be modeled by an
exponential distribution, i.e.,
Pr{Delay > T } = e−ωT
(3)
where Pr{·} denotes the distribution probability, and T represents the delay constraint. ω is the arriving rate, which is
determined by the average delay, i.e.,
ω=
1
.
E{Delay}
(4)
ZHOU et al.: MULTIMEDIA SCHEDULING IN HETEROGENEOUS WIRELESS NETWORKS
E{·} represents the expectation value. Generally, ω needs
to be empirically determined from end-to-end delay statistics
over the network [6]. To present a general solution for online
operation, we construct a model to approximate the average
packet delay.
Let us consider multiple wireless networks N =
{1, 2, . . . , N } simultaneously available to multiple users
S = {1, 2, . . . , S}. Each network n (n ∈ N) is characterized
by its available bit rate ABRn and round-trip time RTTn ,
which are measured and updated periodically. Wireless channel
changes over very short timescales (e.g., up to a few tens of
milliseconds), we assume that ABRn and RTTn represent
average values computed on larger timescales (e.g., one to a
few seconds) and represent the average channel conditions for
user s (s ∈ S) on the given period.
Therefore, the rate allocation can be expressed in a matrix
form, i.e., R = {Rsn }S×N , where the element Rsn represents
the allocated rate of user s (s ∈ S) over network n (n ∈
N). Consequently,
the total allocated rate over network n
= s∈S Rsn , and the total allocated rate for user s is
is Rn Rs = n∈N Rsn . We denote the residual bandwidth (RB) over
network n by RBn . Then, we have
Rsn .
(5)
RBn = ABRn −
For user s in network n, the observed available bandwidth
ABRns is given by
Rsn .
(6)
ABRns = ABRn −
s =s,s ∈S
As the allocated rate on each network approaches the maximum
achievable rate, the average packet delay typically increases
due to network congestion. Similar to [10] and [22], we use
a fractional function to approximate the nonlinear increase in
the packet delay with the traffic rate over network n (n ∈ N),
i.e.,
βn
βn
n
=
n
RBn
ABR −
Rs
(7)
s∈S
where β n can be interpreted as the available source for the
classical M/G/1 queuing model. If we assume equal delay on
both directions, the value of β n can be estimated from the latest
observations of RTTn and RBn , i.e.,
βn =
RBn RTTn
.
2
(8)
If RB is equal to the past observation in network n (n ∈
N) (RBn = RBn ), the average current delay is RTTn /2.
Therefore, for each network n
n n
Rs
2 ABR −
−
Pr{Delay > T } = e−ωT = e
in network n ∈ N is given by
n
= PBn + (1 − PBn ) Pr{Delay > T }
Ploss
n n
Rs
2 ABR −
−
= PBn + (1 − PBn ) e
s∈S
RBn RTTn
T
.
(9)
Together with PBn , which is the random packet-loss rate in
network n due to transmission errors, the total packet-loss rate
s∈S
RBn RTTn
T
.
(10)
The overall distortion from packet loss in network n can be
expressed as
n
n
Dloss
=κPloss
⎛
−
=κ ⎝PBn +(1−PBn )e
2
⎞
ABRn− s∈S Rsn
T
⎠ , n ∈ N.
RBn RTTn
(11)
C. Availability Model
Our availability model is motivated by the reliability models
in the study [29]. Since the availability model is dependent on
the availability cost, we first introduce the availability cost ACns
of user s in network n as follows:
ACns =
s∈S
E{Delay} =
1163
θn
Rsn
(12)
where θn is the unavailable rate of network n, and it can be
expressed as
θn = 1 − e1−(1−ξn )
(13).
ξn (0 ξn 1) is the availability of network n, and is a fixed
system parameter. The value of the parameter must agree with
measurements taken from real systems, whereas the availability
ξn can be estimated and provided by software/hardware vendors. It is worth noting that the way of calculating unavailable
rates is only for illustration purposes, and it is flexible to
substitute any unavailable rate model for (13).
Therefore, the availability cost ACs of user s can be derived
from (12) and (13) as
θn
ACs =
ACnn =
.
(14)
Rsn
n∈N
n∈N
The availability experienced As by user s is expressed as [4]
θn
As = exp[−ACs ] = exp −
.
(15)
Rsn
n∈N
III. D ISTRIBUTED A DAPTIVE S CHEDULING S CHEME
In general, the reconstructed video/audio quality is affected
by both source compression and quality degradation due to
packet losses. We assume that the two distortions are independent and additive. Thus, we can calculate the overall distortion
Dall in terms of mean-square error (MSE) as
Dall = Dcomp + Dloss .
(16)
1164
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 3, MARCH 2011
The objective is to minimize the summation of the total distortion Dall subject to QoS, rate and availability requirements,
i.e.,
min
s∈S,n∈N
Dall (Rsn ) =
s∈S
+
θ
ns
+Ds0
Rs −Rs0
n∈N
κ
2
ABRn− Rsn
−
PBn +(1−PBn )e
n
s∈S
n
RB RTT
T
n∈N
subject to As APQs
ABRns
Rsn = n Rs
ABRs
∀s ∈ S
ϕ(λ) = max ˆlk (λ)
∀n ∈ N
k∈ŝt
n∈N
Rsn ≤ ABRns
∀n ∈ N
g(αk ) = ˆlk (αk ). The approximation function ϕ of g on [α, α ]
can be obtained by connecting these segments.
For any 1 ≤ k ≤ m, we call αk an inflection point if
Ak > Ak+1 . Denote by ατ (1) < ατ (2) < · · · < ατ (q) all inflection points among the breakpoints αk , 1 ≤ k ≤ m; obviously, q ≤ m. Let τ (0) = 0, τ (q + 1) = m + 1, and define
Iˆt = [ατ (t−1) , ατ (t) ] for any a ≤ t ≤ q + 1. It can be seen that
Iˆt is the union of intervals Ik , τ (t − 1) < k < τ (t). Based on
the aforementioned partition, we can obtain a piecewise-convex
expression of the function ϕ, which is very useful to obtain the
global optimization of separable programming problems.
Proposition 1: For any 1 ≤ t ≤ q + 1
(17)
where θs , Rs0 , and Ds0 are the corresponding parameters for
user s ∈ S, and APQs is the availability probability request
by each s. At low rates, the reconstructed quality is limited
by coarse quantization. At high rates, the application stream
will cause more network congestion, which leads to higher
loss rates and reduces the reconstructed quality. For multipleapplication transmission in bandwidth-limited environments,
we therefore expect to achieve maximum decoded quality for
some intermediate rate.
A. Piecewise Approximate Theorem
To obtain the optimal or close-to-optimal result with fast
convergence adapting to the online operation, we propose a
continuous piecewise-linear (CPL) approach to solve the rateallocation optimization based on the utility theory [23]. This
methodology iteratively takes a locally approximate optimal
decision on each user in each network. Our proposed piecewise
approximate method has the following reasoning. The CPL
function is used to approximate the original goal function.
The function is convex in the convex union of many small
hypercubes, and an approximately globally optimal solution of
the original problem confined in this union can be found in
the set of local solutions. In many cases, the number of such
unions may be much less than that of all smaller partitioned
hypercubes. Hence, the CPL approach can substantially reduce
the computation load.
Since Dall is the sum of the univariate functions Dcomp of
each user s and Dloss of each network n, a CPL approximation
can be obtained based on each function by a univariate CPL
function. This can be achieved by partitioning the interest
region of each univariate function into a sufficient number of
nonoverlapping small intervals. Let g be an arbitrary univariate
function whose interest region is [α, α ] ⊂ R. Let m breakpoints α < α1 < α2 < · · · < αm < α be suitably chosen so
that g can be well approximated by the linear function ˆlk =
Ak x + Bk in each small interval Ik = [αk−1 , αk ] for any 1 ≤
k ≤ m + 1, where α0 = α, αm+1 = α , and Ak and Bk are
determined by the linear equations g(αk−1 ) = ˆlk (αk−1 ) and
∀λ ∈ Iˆt
(18)
where ŝt = k, τ (t − 1) < k ≤ τ (t).
Proof: For any 1 ≤ t ≤ q + 1, since Aτ (t−1)+1 <
Aτ (t−1)+2 < · · · < Aτ (t) , function ϕ is convex on Iˆt for any
arbitrarily chosen λ̂ ∈ Iˆt [24]. There should be an integer
k ∈ ŝt such that λ̂ ∈ Ik and ϕ(λ̂) = ˆlk (λ̂). Then, the following
inequality is satisfied:
ˆlk (λ̂) ≤ ˆlk (λ̂)
∀k ∈ ŝt − {k}.
(19)
Otherwise, there could be k (k ∈ ŝt − {k}) such that ˆlk (λ̂) >
ˆlk (λ̂). We could then choose λ̄ ∈ Iˆk ⊂ Iˆt and a sufficiently
small positive number ε such that λ = ελ̂ + (1 − ε)λ̄ ∈ Ik .
Because ϕ is convex on Iˆt , we have
ϕ(λ ) ≤ εϕ(λ̂) + (1 − ε)ϕ(λ̄) = εˆlk (λ̂) + (1 − ε)ˆlk (λ̄)
< εˆlk (λ̂) + (1 − ε)ˆlk (λ̄) = ˆlk (λ )
(20)
which contradicts the known relation ϕ(λ ) = ˆlk (λ ). As λ̂ is
arbitrarily chosen, the proposition is proved.
According to the aforementioned proposition, we can partition the original interest region into a number of smaller
nonoverlapping hypercubes and approximate the goal function
on every hypercube by a convex CPL function. In fact, the goal
function Dall corresponds to the arbitrary univariate function g,
and any potential rate allocation is the breakpoint of the aforementioned proposition. Therefore, how to find the appropriate
breakpoints and judge whether it is an inflection point is the key
point for implementing the piecewise approximate proposition
of our rate-allocation problem. To get around this difficulty, we
employ a utility-based function in the proposed utility-based
rate-allocation (UBRA) algorithm.
B. Rate-Allocation Algorithm
To fully satisfy the user’s availability requirements, the
proposed UBRA tends to assign applications to a group of
networks that can provide high availability levels. Note that
the availability level offered by a network is proportional to its
transmission ability. This implies that UBRA might assign a
large number of applications into a network with high availability level and transmission ability. As a result, the mean
ZHOU et al.: MULTIMEDIA SCHEDULING IN HETEROGENEOUS WIRELESS NETWORKS
QoS achieved by UBRA could significantly suffer from a load
imbalance. To prevent severe load imbalance, UBRA leverages
a load-imbalance detection mechanism, which is called a loadimbalance detector (LID), to detect whether a network n in the
heterogeneous networks is overloaded. The LID uses a load
index Ln to measure relative network n workload, as defined
in the following equation:
ACns
s∈S
(21)
Ln = n
s∈S
n∈N ACs /N
where s∈S ACns is the availability cost of network n, and
n
s∈S
n∈N ACs /N is the average network cost of the whole
system. When Ln is higher than a threshold value TL, network
n is considered as overloaded. Note that TL is an empirical
parameter.
We define Rsn → Rsn as the transition of the next allocation
rate for the user s ∈ S in network n ∈ N, and Rsn is selected in
the set of RAi (i = 1, 2). Rsn = Rsn + Rsn , where Rsn is the
rate improvement at each iteration.1 The utility of this transition
can be computed as [10]
Usn =
ϕ(Rsn ) − ϕ(Rsn )
Rsn − Rsn
(22)
where ϕ is the approximate linear function for Dall in the interval [Rsn , Rsn ]. The total utility matrix is U = {Usn }S×N . During
each iteration, the proposed algorithm finds R = {Rsn }S×N
that brings the highest utility U∗ = {Usn }S×N to the overall
system by its transition, i.e.,
U∗ = arg max U.
R
(23)
One starts to allocate resources to user s in network n. Once
the resources of network n are depleted, the algorithm will find
a different user that can free the required resources for user s
in another network by allocating part of its rate. This operation
is performed as long as the overall utility of the system can be
improved and as long as free network resources still exist in
the overall system. The algorithm stops when there are no more
free resources in the network system or when no other possible
user transition can improve the overall system utility.
Algorithm 1 UBRA algorithm
1. Input:
2. PBn , ABRns , RTTn , θn , ξn , APQs ∀ user s ∈ S in network
n ∈ N;
3. Rsn = 0, Rsn = ABRns /2, ∀ user s ∈ S in network n ∈
N;
4. Output:
5. Global Rate Allocation R;
6. Procedure RateAllocation
7. while (true)
8. for s = 1 to S do
1 In theory, the initial Rn can be chosen at random as long as it is less than
s
n
n
ABRn
s ; here, we set Rs to ABRs /2 as the initial value.
1165
9. for n = 1 to N do
10.
compute ACns and Ln ;
11.
compute the utility of Rsn → Rsn : Usn = ϕ(Rsn ) −
ϕ(Rsn )/(Rsn − Rsn );
12.
Rsn = Rsn /Usn ;
Rsn = Rsn + Rsn ;
13.
14.
update the approximate function ϕ;
15. end for
16. end for
17. find U∗ = arg maxR U;
18. IntraNet(R, U∗ , n);
19. Procedure IntraNet(R, U∗ , n)
20. if Ln ≤ TL then
21. Rsn → Rsn ;
22. update free resources on network n;
23. else
24. InterNet(R, U∗ , n);
25. end if
26. Procedure InterNet(R, U∗ , n)
27. find other user that can transfer part of its allocated to
network n = n ∈ N with maximum transition utility
improvement U;
28. if U > 0 then
29. perform the resource free up:
30. Rsn → Rsn ;
31. update free resources on network n and n ;
32. else
33. break;
34. end if
Algorithm 1 represents a sketch of the proposed UBRA. In
this algorithm, the IntraNet always attempts to increase the
system’s utility by allocating some resource in network n (n ∈
N). If the available resources are not sufficient, the InterNet
procedure will find a new user that can free up enough resources
by allocating parts of its allocated rate through another network
n = n ∈ N. As long as there remain available network resources, the procedures go on until no extra utility improvement
can be brought to the overall system [6].
Theorem 1: The time complexity in the worst case of UBRA
is O(N × S), where N is the number of networks, and S is the
number of users.
Proof: The complexity involved in the search operations
for optimal networks is O(N ). This is also true for the IntraNet
procedure. In the worst case, the algorithm requires O(S) iterations to pass through every user. Hence, the total complexity of
the algorithm is O(N × S).
Since N and S in practice are all finite integers, Theorem 1
shows that the time complexity of UBRA is acceptably low in
most cases. This time complexity indicates that the execution
time of UBRA is lower compared with the application transmission. Thus, the overhead for executing UBRA can be reasonably
ignored in our experiments.
Theorem 2: In a workload where the maximal availability
requirement among all users is less than or equal to the minimal
availability among all networks, the availability cost of user
s ACs is zero.
1166
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 3, MARCH 2011
Proof: Since the maximal availability requirement among
all users is not more than the minimal availability among all
networks, the unavailable
rate of network n is zero, i.e., θn = 0.
Therefore, ACs = n∈N (θn /Rsn ) = 0.
Theorem 2 demonstrates that, if each network n can fully
satisfy the availability requirements of tasks for any user s,
there is no availability cost in network n for each user s.
Theorem 3: Given user s whose availability requirement
cannot be satisfied by any node in a heterogeneous system with
N networks, UBRA initially allocates user s to a network n
whose availability is the highest among all networks. In this
case, the whole system achieves minimum distortion.
Proof: Let AN be the availability of all the N networks,
with the elements in increasing order of their availability. A
valid rate allocation should satisfy the utility constraints. Let
An be the network with the largest availability. If the rate
in network n is not the largest one, one can always find a
better solution by transferring the rate from other networks
sharing the same network resources. Since the total rate stays
constant at that moment, the rate transfer does not violate the
multiple network constraints. It, however, changes the total
source distortion, resulting in an increased distortion.
Theorem 3 establishes that, if all the networks in a heterogeneous system are incapable of guaranteeing the availability
requirements of a user, UBRA initially allocates the user to a
network with the highest availability among all the networks in
the system.
To adapt source rates at the transport layer according to
the network states reported from the network layer, crosslayer information exchange is necessary. Specifically, at the
network layer, the distributed allocation scheme needs to track
the observations of ABRn and RTTn over all available access
networks. It also records the required rate Rsn by each user
and then calculates the values of Dall and Usn . At the transport
layer, the rate controller at the source advertises its intended
rate allocation Rsn . The network state monitor traversed by
the stream then calculates the relevant parameters based on
its local cache of ABRn , RTTn , and RBn within the same
access network. The destination node extracts such information
from the packet header and reports back to the sender with
some acknowledgment packets so that the rate controller can
reoptimize its intended rate Rsn based on the proposed UBRA,
with updated network state information.
IV. N UMERICAL R ESULTS
Here, we conduct simulations to evaluate the performance of
the proposed rate-allocation scheme in heterogeneous wireless
networks. First, we describe the simulation environment and
key parameters. Then, we present the simulation results in
different scenarios.
A. Simulation Environment
To simulate video and audio applications, we use the highdefinition (HD) City video sequence and the broadcasting audio
sequence as the test sequences in our simulations. For the audio
application, we use four audio transcoders, namely, G.723.1B,
TABLE I
S TATISTICS OF M EASURED ABR AND RTT
iLBC, SPEEX, and G.711, with average encoding rates of 6.4,
15.2, 24.6, and 64 kb/s, respectively. In terms of HD video,
the sequence has a spatial resolution of 1280 × 720 pixels and
a frame rate of F = 60 frames/s. Video streams are encoded
using a fast implementation of the H.264/AVC codec [25] at
various quantization step sizes, with a group-of-pictures (GOP)
length of 30 and an IBBP structure similar to that often used in
Moving Pictures Experts Group-2 bitstreams. Encoded video
frames are segmented into packets with maximum size of
1500 bytes, The transmission intervals of each packet within the
entire GOP are spread out evenly, which results in unnecessary
queuing delay due to the large sizes of intracoded frames. For
the sake of simplicity, we use the constant bit rate (CBR) model
to represent audio/video traffic.
We allow each user to stream video or audio applications via
two wireless networks (802.11b and 802.11g) with a maximum
allowable total delay T = 350 ms. We choose the availability
parameters = 0.1, APQs = 0.9, and TL = 1.2. ξn is randomly chosen between 0.8 and 1.0. Each network is simulated
as a link with varying available bandwidth and delay according
to the traces collected from the actual access networks using the
ABR and RTT measurements.2 Table I summarizes the statistics
of the collected ABR and RTT of each network tracing over
200 GOP periods. During transmission, the environments are
updated every frame transmission, which can cause changes in
the rate allocation and network resources. During successive
frame transmission intervals, the environment is unchanged.
It should be noted that all the simulation results have been
averaged over 300 runs.
We also present an example of a quality metric based on the
mean opinion score (MOS) value. The MOS captures the average user satisfaction on a scale stemming from 1 to 4.5.3 The
minimum value reflects an unacceptable application quality,
whereas the maximum value demonstrates an excellent QoS.
Traditionally, the method of determining audio quality is the
mode of Perceptual Evaluation of Speech Quality (PESQ) [26]
standardized by the International Telecommunication Union.
However, the PESQ algorithm is computationally too expensive
to be used in real-time scenarios. To address this issue, we
employ a model that maps the quality of audio streams into
MOS values. Therefore, the perceived quality of each of the two
applications can be converted into an equivalent MOS, which
is later used in the optimization problem. The performance
of different audio transcoders as a function of network losses
is mapped to MOS values using the relationship between the
rate and the packet-loss rate. Fig. 2 shows the MOS estimation
2 Forward and backward trip delays are both simulated as half of the measured RTT.
3 For our simulations, each MOS value is averaged over 100 experiments.
ZHOU et al.: MULTIMEDIA SCHEDULING IN HETEROGENEOUS WIRELESS NETWORKS
Fig. 2.
MOS versus packet-loss rate under different transmission rates.
Fig. 3.
Video PSNR–MOS mapping.
in terms of packet-loss rate with different audio codecs [27].
The curves are drawn using an average over a large number of
audio samples and channel realizations (packet-loss patterns).
These curves can be stored in the server for every codec. The
perceived video-streaming quality is initially mapped into a
peak signal-to-noise ratio (PSNR) distortion measure. Fig. 3
shows the mapping between the PSNR and the MOS [27].
B. Performance Evaluation of the Proposed Scheme
In order validate the distortion model introduced in
Section II-B, Fig. 4 shows the rate–PSNR tradeoff when one
user streams the HD video sequence City (300 frames) over two
wireless networks, i.e., 802.11b and 802.11g, respectively. The
analytical results match the experimental data very well. In the
first case, the only losses due to late arrivals are considered. In
the second case, an additional end-to-end random loss rate of
5% is taken into account. The bell shape of the curves illustrates
that the highest performance is obtained when the streaming
rate achieves the optimal tradeoff between compression quality
and self-inflicted congestion.
To demonstrate the effectiveness of our proposed UBRA
algorithm, we use the representative drop-tail scheme, which
1167
employs the fixed-rate allocation method, and the additiveincrease–multiplicative-decrease (AIMD)-based rate-allocation
method, which is used by TCP congestion control [28], for
comparison. To have a clear understanding of the reconstructed
video quality, we only consider one video over an 802.11b
network. The drop-tail scheme employs a fixed-source coding rate Rf = 1.50 Mb/s. When the rate exceeds the optimal
transmission rate for the selected source–destination pair, it
will drop the subsequent encoded packets. The AIMD-based
scheme probes the network for available bandwidth and reduces
the rate allocation after congestion occurs. Each user s initiates
its rate at a specified rate RsAIMD , corresponding to the minimum acceptable video quality, and increases its allocation by
Rs every t seconds. In the case of network congestion,
the allocated rate is dropped by (Rsn − RsAIM D )/2 over the
congested network n. The increase in rate allocation is allocated
to all available networks in proportion to the average ABR
of each. In addition, congestion over network n is indicated
upon detection of the lost packets, or when the observed RTT
exceeds a specified threshold, based on the playout deadline of
the audio/video stream.
Fig. 5 shows the comparison among our proposed rateallocation scheme, the drop-tail scheme, and AIMD. In the scenario, packet losses are caused only by channel overpumping.4
It should be noted that, due to the use of CBR encoding, the
video quality is not constant and varies periodically [5]. In
Fig. 5, the average PSNR using the proposed UBRA is 35.3 dB,
whereas it is 35.0 using the AIMD-based method and 34.7 dB
in the case of drop tail. Thus, the proposed UBRA can achieve
almost 0.3- and 0.60dB performance gains compared with the
AIMD and drop-tail schemes, respectively. The average rates
computed by the proposed scheme for the four GOPs are 1.54,
2.02, 1.21, and 1.86 Mb/s, respectively. We can see that for
GOP No. 1, No. 2, and No. 4, the allocated transmission rate
using the proposed UBRA method is higher than the fixed
1.50 Mb/s. Thus, our rate-allocation proposal allows for the
improvement of the performance compared with using a fixedrate coding scheme. On the other hand, for GOP No. 3, it
is clear that the fixed-source coding rate is higher than the
allocated transmission rate. Therefore, packet losses will occur
when the transmission buffer is full. This will result in frames
loss and, hence, cause substantial performance degradation. A
lost frame is concealed by just copying the previous frame.
If several consecutive frames are lost, the degradation will be
even more serious since the concealed frames are then used as
correctly received frames to conceal the subsequent lost frames.
This results in substantial error propagation. For example, in
Fig. 5, we can see that there is a substantial performance
degradation around the 90th frame for the no-rate control
case due to channel overpumping. Furthermore, although the
performance degradation caused by the channel overpumping
packet losses has been partially compensated using the passiveerror-concealment technique, the performance is still not as
good as using the rate-allocation scheme. However, Fig. 5
shows that the proposed UBRA still outperforms the traditional
4 Here,
we assume that no transmission errors occurred.
1168
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 3, MARCH 2011
Fig. 4. Decoded video quality approximate model and experimental data. The value of κ is 395 for both networks. (a) 802.11b (ABR = 4.5 Mb/s, RTT =
221 ms). (b) 802.11g (ABR = 16.1 Mb/s, RTT = 290 ms).
comes from the fact that the video applications usually require
more network resources than audio applications.
V. C ONCLUSION
Fig. 5. Performance comparison between different rate-allocation schemes.
AIMD-based method. On one hand, although the AIMD-based
method can adapt itself to the network condition, the network
is so dynamic that a congested node forwarding a few seconds might not be used at all when the source reacts to the
congestion. On the other hand, the proposed UBRA method
further takes advantage of the explicit knowledge of the video
distortion-rate characteristics and can achieve balanced video
quality.
It is very important to study the network performance and
video transmission in dynamic networks with user dynamics,
e.g., join and leave. We start with four users (two video applications and two audio applications). At time t = 2 s, we add four
users (two video applications and two audio applications), and
at time t = 8 s, we randomly remove two users (one video application and one audio application). Fig. 6(a) and (b) shows the
average MOS for the original four users that are obtained by the
proposed UBRA method and the AIMD method, respectively.
We observe that our proposal outperforms the AIMD method
on the aspect of constant performance. This occurs because
UBRA manages to keep a rather constant application quality for
all active users by redistributing parts of the network resources
to the new users. Fig. 6(a) and (b) also shows that the audio
applications are less sensitive than the video applications. This
In this paper, we have first developed and evaluated a
framework for availability-aware multimedia applications over
multiple heterogeneous wireless networks based on observing
each access network parameter, as well as each application
rate-distortion and availability characteristic. Then, we have
proposed an availability-aware adaptive scheduling algorithm
for online operation, and this algorithm achieves an optimal
or close-to-optimal end-to-end QoS under the overall limited
resource budget. The simulation results demonstrate the effectiveness of our proposed rate-allocation scheme for multiapplication service in heterogeneous wireless networks.
We believe that these results offer new insights into the
design of efficient scheduling policies for video streaming over
heterogeneous wireless networks. First, they provide additional
insights that motivate the study of relevant protocols of multimedia transmission over heterogeneous networks. One could
argue that we can use existing protocols directly to the
availability-aware case for convenience. The results of this
paper indicate that this is not always the case. Second, a main
difference between the protocol developed in this paper and the
current IMS platform is that the latter takes up a few milliseconds to switch networks, whereas our work assumes that the
networks can be switched dynamically. Therefore, we believe
that our results provide a strong motivation to purse improved
network-switching hardware and protocol in the future. Third,
the main advantages of the proposed availability-aware strategy
result from the prior knowledge of the source media requirements; however, obtaining this kind of information may be very
difficult in some applications (i.e., live show, etc.). Therefore,
this paper motivates the study of media-aware scheduling with
blind or simplified source availability information.
For future work, we plan to study some practical issues for
implementing the proposed schemes. Note that, in real multimedia transmission over heterogeneous wireless networks,
ZHOU et al.: MULTIMEDIA SCHEDULING IN HETEROGENEOUS WIRELESS NETWORKS
Fig. 6.
1169
Comparison with average performance per user in case users join/leave the network. (a) UBRA. (b) AIMD.
additional works need to be developed to 1) reduce the dependence of the media content or scheduling scheme to automatically adapt the original media content; 2) simplify the adaptive
scheduling scheme, particularly the network and source information exchange and feedback; and 3) extend the results to
more practical systems (e.g., multichannel multiradio wireless
networks). In our ongoing work, we plan to carefully address
these open problems and study their impact on the actual
heterogeneous systems.
R EFERENCES
[1] P. Vidales, J. Baliosion, J. Serrat, G. Mapp, F. Stejano, and A. Hopper,
“Autonomic system for mobility support in 4G networks,” IEEE J. Sel.
Areas Commun., vol. 23, no. 12, pp. 2288–2304, Dec. 2005.
[2] A. Cuevas, J. I. Moreno, P. Vidales, and H. Einsiedler, “The IMS platform: A solution for next generation network operators to be more than
bit pipes,” IEEE Commun. Mag.—Special Issue on Advances of Service
Platform Technologies, vol. 44, no. 8, pp. 75–81, Aug. 2006.
[3] L. Zhou, N. Xiong, L. Shu, A. Vasilakos, and S.-S. Yeo, “Contextaware multimedia service in heterogeneous networks,” IEEE Intell. Syst.,
vol. 25, no. 2, pp. 40–47, Mar./Apr. 2010.
[4] X. Qin and T. Xie, “An availability-aware task scheduling strategy for
heterogeneous systems,” IEEE Trans. Comput., vol. 57, no. 2, pp. 188–
199, Feb. 2008.
[5] Q. Qu, Y. Pei, and J. W. Modestino, “An adaptive motion-based unequal error protection approach for real-time video transport over wireless
IP networks,” IEEE Trans. Multimedia, vol. 8, no. 5, pp. 1033–1044,
Oct. 2006.
[6] L. Zhou, B. Geller, A. Wei, and B. Zheng, “Cross-layer rate allocation for
multimedia applications in pervasive computing environment,” in Proc.
IEEE GLOBECOM, 2008, pp. 1–5.
[7] H. P. Shiang and M. van der Schaar, “Queuing-based dynamic channel
selection for heterogeneous multimedia applications over cognitive radio networks,” IEEE Trans. Multimedia, vol. 10, no. 5, pp. 896–909,
Aug. 2008.
[8] M. van der Schaar and Y. Andreopoulos, “Rate-distortion-complexity
modeling for network and receiver aware adaptation,” IEEE Trans. Multimedia, vol. 7, no. 3, pp. 471–479, Jun. 2005.
[9] L. Zhou, X. Wang, W. Tu, G. Muntean, and B. Geller, “Distributed
scheduling scheme for video streaming over multi-channel multi-radio
multi-hop wireless networks,” IEEE J. Sel. Areas Commun., vol. 28, no. 3,
pp. 409–419, Apr. 2010.
[10] L. Zhou, B. Geller, X. Wang, A. Wei, B. Zheng, and H.-C. Chao, “Multiuser video streaming over multiple heterogeneous wireless networks: A
distributed, cross-layer design paradigm,” J. Internet Technol., vol. 10,
no. 1, pp. 1–12, 2009.
[11] D. Wu, S. Ci, H. Luo, H. Wang, and A. Katsaggelos, “A quality-driven decision engine for service-oriented live video transmission,” IEEE Wireless
Commun., vol. 16, no. 4, pp. 48–54, Aug. 2009.
[12] A. H. Zahran and B. Liang, “A generic framework for mobility modeling and performance analysis in next-generation heterogeneous wireless
networks,” IEEE Commun. Mag., vol. 45, no. 9, pp. 92–99, Sep. 2007.
[13] Y. Gou, D. A. J. Pearce, and P. D. Mitchell, “A receiver-based vertical
handover mechanism for TCP congestion control,” IEEE Trans. Wireless
Commun., vol. 5, no. 10, pp. 2824–2833, Oct. 2006.
[14] M. Handley, S. Floyd, J. Pahdye, and J. Widmer, TCP Friendly Rate
Control (TFRC): Protocol Spec., Jan. 2003, RFC 3448.
[15] J. Yan, K. Katrinis, M. May, and B. Plattner, “Media- and TCP-friendly
congestion control for scalable video streams,” IEEE Trans. Multimedia,
vol. 8, no. 2, pp. 196–206, Apr. 2006.
[16] T. Alpcan and T. Basar, “A utility-based congestion control scheme for
Internet-style networks with delay,” IEEE/ACM Trans. Netw., vol. 13,
no. 6, pp. 1261–1274, Dec. 2005.
[17] H. Chen and Z. Liu, “Extending TCP congestion control to multicast,”
Comput. Netw., vol. 51, no. 11, pp. 3090–3109, Aug. 2007.
[18] A. Szwabe, A. Schorr, F. J. Hauck, and A. J. Kassler, “Dynamic multimedia stream adaptation and rate control for heterogeneous networks,”
in Proc. 15th Int. Packet Video Workshop, Hangzhou, China, May 2006,
vol. 7, pp. 63–69.
[19] W. Tu, J. Chakareski, and E. Steinbach, “Rate-distortion optimized
frame dropping for multi-user streaming and conversational video,” Adv.
Multimedia—Special Issue on Collaboration and Optimization for Multimedia Communications, vol. 2008, pp. 1–13, 2008, Article ID 628970,
DOI:10.1155/2008/628970.
[20] K. Stuhlmuller, N. Farber, M. Link, and B. Girod, “Analysis of video
transmission over lossy channels,” IEEE J. Sel. Areas Commun., vol. 18,
no. 6, pp. 1012–1032, Jun. 2000.
[21] X. Zhu, E. Setton, and B. Girod, “Congestion-distortion optimized video
transmission over ad hoc networks,” Signal Process.: Image Commun.,
vol. 20, no. 8, pp. 773–783, Sep. 2005.
[22] X. Zhu, P. Agrawal, J. P. Singh, T. Alpcan, and B. Girod, “Rate allocation
for multi-user video streaming over heterogeneous access networks,” in
Proc. ACM Multimedia, Augsburg, Germany, Sep. 2007, pp. 37–46.
[23] F. Kelly and T. Voice, “Stability of end-to-end algorithm for joint routing
and rate control,” ACM SIGCOMM Comput. Commun. Rev., vol. 32, no. 2,
pp. 5–12, Apr. 2005.
[24] G. B. Dantzig, Linear Programming and Extensions. Princeton, NJ:
Princeton Univ. Press, 1963.
[25] Advanced Video Coding for Generic Audiovisual Services, ITU-T Recommendation H.264-ISO/IEC 14496-10 (AVC), 2003.
[26] PESQ: An Objective Method for End-to-End Speech Quality Assessment
of Narrow-Band Telephone Networks and Speech Codecs, ITU-T P.862,
2001.
[27] D. Jurca, W. Kellerer, E. Steinbach, S. Khan, S. Thakolsri, and P. Frossard,
“Joint network and rate allocation for video streaming over multiple wireless networks,” in Proc. IEEE Int. Symp. Multimedia, Taichung, Taiwan,
Dec. 10–12, 2007, pp. 229–236.
[28] E. Altman and K. Avrachenkov, “Performance analysis of AIMD mechanisms over a multi-state Markovian path,” Comput. Netw., vol. 47, no. 3,
pp. 307–326, Feb. 2005.
[29] S. Srinivasn and N. K. Jha, “Safety and reliability driven task allocation
in distributed systems,” IEEE Trans. Parallel Distrib. Syst., vol. 10, no. 3,
pp. 238–251, Mar. 1999.
1170
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 60, NO. 3, MARCH 2011
Liang Zhou received the Ph.D. degree in electrical
engineering from both the Ecole Normale Supérieure
(E.N.S.), Cachan, France, and Shanghai Jiao Tong
University, Shanghai, China, in March 2009.
From 2009 to 2010, he was a postdoctoral
researcher with ENSTA-ParisTech, Paris, France.
Since October 2010, he has been an Alexander von
Humboldt Research Fellow with Munich University
of Technology, Munich, Germany. His research interests are in the area of network-based multimedia communications and computing, in particular,
resource allocation, system scheduling, cross-layer design, multimedia security,
and multimedia signal processing.
Yan Zhang received the Ph.D. degree from Nanyang
Technological University, Singapore.
Since August 2006, he has been with Simula
Research Laboratory, Norway, where he is currently
heading the “Wireless Networks” research group. He
is also an Associate Professor (part-time) with the
University of Oslo, Norway. He is a regional editor,
associate editor, on the editorial board, or guest editor of a number of international journals. He serves
as organizing committee chair for many international
conferences. His research interests include resource,
mobility, spectrum, energy, and data management in wireless communications
and networking.
Haohong Wang received the B.S. degree in computer science and the M.Eng. degree in computer applications from Nanjing University, Nanjing, China,
the M.S. degree in computer science from the University of New Mexico, Albuquerque, and the Ph.D.
degree in electrical and computer engineering from
Northwestern University, Evanston, IL.
He held various technical and management positions at AT&T, Catapult Communications, Qualcomm, Marvell, and TCL-Thomson Electronics. He
is currently with Cisco Systems, San Jose, CA, where
he is leading efforts on Telepresence initiatives. His research involves the areas
of multimedia communications, 3-D graphics, video analysis and processing,
and 3-D television systems.
Athanasios V. Vasilakos is a Professor with the
Department of Computer and Telecommunications
Engineering, University of Western Macedonia,
Greece, and a Visiting Professor with the Graduate Programme, Department of Electrical and Computer Engineering, National Technical University of
Athens, Athens, Greece. He was or is on the editorial boards of more than 20 international journals, including the IEEE Communications Magazine,
(1999–2002 and 2008-), the IEEE T RANSACTIONS
ON S YSTEMS , M AN , AND C YBERNETICS , PART B
(2007-), the IEEE T RANSACTIONS ON I NFORMATION T ECHNOLOGY IN
B IOMEDICINE (2008-), the IEEE T RANSACTIONS ON W IRELESS C OMMU NICATIONS , (invited), and the ACM Transactions on Autonomous and Adaptive
Systems, (invited). He is chairman of the Telecommunications Task Force
of the Intelligent Systems Applications Technical Committee of the IEEE
Computational Intelligence Society.
Shiguo Lian received the Ph.D. degree in multimedia security from Nanjing University of Science and
Technology, Nanjing, China, in 2005.
In 2004, he was a Research Assistant with the
City University of Hong Kong. Since July 2005,
he has been with France Telecom Research and
Development (Orange Laboratories), Beijing, China.
His research interests include network and multimedia security and intelligent services, i.e., lightweight
cryptography, digital rights management, and intelligent multimedia services and security.
Weiping Jing is a researcher with the Jiangsu
Province Key Laboratory of ASIC Design, Nantong
University, Nantong, China.
His researcher interests include vehicular networks, intelligent transportation, vehicular communications, etc.