Joint Optimization of User-Experience and Energy

1522
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 7, JULY 2014
Joint Optimization of User-Experience and
Energy-Efficiency in Wireless Multimedia
Broadcast
Chetna Singhal, Student Member, IEEE, Swades De, Member, IEEE, Ramona Trestian, Member, IEEE,
and Gabriel-Miro Muntean, Member, IEEE
Abstract—This paper presents a novel cross-layer optimization framework to improve the quality of user experience (QoE) and
energy efficiency of the heterogeneous wireless multimedia broadcast receivers. This joint optimization is achieved by grouping the
users based on their device capabilities and estimated channel conditions experienced by them and broadcasting adaptive content to
these groups. The adaptive multimedia content is obtained by using scalable video coding (SVC) with optimal source encoding
parameters resulted from an innovative cooperative game. Energy saving at user terminals results from using a layer-aware time
slicing approach in the transmission stage. A trade-off between energy saving and QoE is observed, and is incorporated in the
definition of a utility function of the players in the formulated heterogeneous user composition and physical channel aware game. An
adaptive modulation and coding scheme is also optimally incorporated in order to maximize the reception quality of the broadcast
receivers, while maximizing the network broadcast capacity. Compared to the conventional broadcast schemes, the proposed
framework shows an appreciable improvement in QoE levels for all users, while achieving higher energy-savings for the energy
constrained users.
Index Terms—Adaptive multimedia broadcast and multicast, scalable video coding, adaptive modulation and coding, heterogeneous
users, energy saving, quality of user experience
1
I NTRODUCTION
R
APID advancement in communication technologies in
recent years, coupled with the availability of affordable
high-end mobile computing devices, such as smartphones,
tablets, personal digital assistants, small notebooks, have
led to a significant growth in the number of consumers that
access multimedia services from various types of devices,
while on the move or stationary [1], [2].
The prevalent wireless technologies for multimedia
broadcast include Long Term Evolution (LTE) using
extended Multimedia Broadcast and Multicast Services
(e-MBMS) interface specifications [3], [4], Worldwide
Interoperability for Microwave Access (WiMAX) [3], and
Digital Video Broadcast (DVB) [5]–[7]. Although the latest
advances in many wireless network technologies, including
broadcast (e.g. DVB-second generation terrestrial (DVB-T2),
DVB-hand-held (DVB-H)), broadband (e.g. IEEE 802.11g,
• C. Singhal is with the Bharti School of Telecommunications, Indian
Institute of Technology (IIT) Delhi, New Delhi 110016, India.
E-mail: [email protected]
• S. De is with the Electrical Engineering Department, IIT Delhi, New
Delhi 110016, India. E-mail: [email protected]
• R. Trestian is with the Computer Communication Department, Middlesex
University, London NW4 4BT, U.K. E-mail: [email protected]
• G.-M. Muntean is with the Performance Engineering Laboratory, School
of Electronic Engineering, Dublin City University, Dublin 9, Ireland.
E-mail: [email protected]
Manuscript received 18 June 2013; revised 7 Oct. 2013; accepted 19 Oct.
2013. Date of publication 23 Oct. 2013; date of current version 2 July 2014.
For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference the Digital Object Identifier below.
Digital Object Identifier 10.1109/TMC.2013.138
IEEE 802.11n [8]), and cellular (e.g. LTE), have enabled the
operators to increase network capacity, the demands for
popular multimedia content delivery to the mobile devices
are growing even faster. Consequently, the overall user
experience is still far from optimal, as the rich multimedia content puts pressure on the existing communication
resources in terms of their bandwidth requirements and
real-time constraints.
Thus, the challenge for the network operators include
network resource optimization for popular multimedia content delivery, while ensuring uninterrupted and smooth
services over wireless to a diverse customer population
with varying degree of user-end constraints.
1.1 Motivation and Proposed Solution
In multimedia broadcast, one challenge is posed by userend heterogeneity (e.g., different display size, processing
capabilities, channel impairments). Another key component
that consumers highly care about is the battery lifetime of
their high-end mobile device. It is known that, real-time
multimedia applications demand strict Quality of Service
(QoS), but they are also very power-hungry.
Given the above user-end constraints, a service provider
would look for maximizing the number of users served
without affecting the Quality of user Experience (QoE).
Clearly, attempting to receive a broadcast content irrespective of the device constraints is detrimental to battery
resource efficiency, wherein the low-resolution mobile users
suffer from redundant processing of high-end data that the
device is not even able to use fully.
c 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
1536-1233 See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
SINGHAL ET AL.: JOINT OPTIMIZATION OF USER-EXPERIENCE AND ENERGY-EFFICIENCY IN WIRELESS MULTIMEDIA BROADCAST
There have been a few recent studies that address
receiver energy constraints [9], [10], display limitations
and channel dynamics [11]–[14], source and channel rate
adaptation [15]. Yet to our best knowledge, a comprehensive look into the optimal broadcast strategy that jointly
caters to both user-specific constraints and network dynamics is still missing.
This paper presents a novel cross-layer optimization framework to improve both user QoE levels and energy efficiency of
wireless multimedia broadcast receivers with varying display and energy constraints. This solution combines user
composition-aware source coding rate (SVC) optimization,
optimum time slicing for layer coded transmission, and a
cross-layer adaptive modulation and coding scheme (MCS).
1.2 Key Features and Findings
The main features of the proposed framework are as follows: 1) user grouping based on individual device capabilities and channel conditions; 2) formulation of a cooperative
game to obtain user heterogeneity aware optimized SVC
parameters that enable energy saving of the battery constrained users and at the same time maintain high QoE
levels for high-end users; 3) optimizing layer-coded time
slicing for energy saving and quality trade-off; 4) user heterogeneity and physical channel adaptive MCS allocation to
the layered video content that maximizes network capacity.
The main findings of this work are: (a) The proposed
user- and channel-aware grouping and cooperative game
provide the users options to trade between quality of
reception and energy conservation. (b) the usage of time
slicing along with user heterogeneity and channel aware
MCS significantly reduce energy consumption and increase
QoE; the number of users served in the network with a
guaranteed minimum quality level is increased.
Specifically, tests in different traffic scenarios reveal that,
the proposed adaptive MCS offers about 16.6% higher
user serving capacity compared to fixed MCS or simple
MCS schemes. With respect to only energy saving based
optimization, the proposed joint energy and quality based
cross-layer optimizations give about 43% higher video quality, while trading off only about 8% in energy saving and a
marginal 0.62% in user serving capacity. Compared to only
quality based optimization, the proposed scheme results in
about 17% extra energy saving, 3.5% higher quality, and
10.8% higher capacity.
1.3 Paper Organization
The rest of this paper is organized as follows: Section 2
discusses related works and Section 3 presents the technological details of the system and the proposed framework.
This is followed by the analytic system performance model
and optimizations in Section 4. Subsequently, Section 5
describes the simulation framework and Section 6 presents
the key results of the proposed user-centric optimized multimedia broadcast scheme. Finally, the paper is concluded
in Section 7.
2
R ELATED W ORKS
Hierarchical video coding [16] is an attractive solution
that allows a user to dynamically adapt the video bitstream reception in dynamic wireless channel conditions.
1523
This technique encodes the stream into multiple progressively dependent layers. The most important layer is called
base layer which typically provides an acceptable basic
quality. The rest of the layers are known as enhancement
layers which can be added to the base layer to improve
the video quality. To this end, both ITU-T VCEG and
ISO/IEC MPEG have standardized the SVC [17], [18] extension of H.264/AVC [19]–[21]. The H.264/SVC extension
achieves a rate-distortion performance comparable to that
of H.264/AVC, where the same visual perceived quality is
typically achieved with at most 10% higher bit rate [22].
DVB-H, an European Telecommunications Standards
Institute (ETSI) standard [23], provides a built-in function that helps exploiting the video scalability features
using Hierarchical Modulation [24] and is an efficient way
to broadcast multimedia services over digital terrestrial
networks to hand-held terminals. However, it considers
transmission level details only, but not the user constraints
or video encoding details.
[25] compared group management mechanisms in IP
and MBMS models in UMTS networks, but did not discuss
group formation criterion and user heterogeneity. An adaptive radio resource allocation scheme for multi-resolution
multicast services in orthogonal frequency-division multiplexing (OFDM) systems was proposed in [26], which was
shown to achieve an improved system throughput while
maintaining fairness among all users. For energy-efficient
streaming of scalable video over LTE using e-MBMS, grouping of users based on position and requested video quality
was considered in [11]. Discontinuous reception (DRX) and
energy saving at the user-end was not considered here;
instead energy saving at the base station (BS) was targeted.
A cross-layer adaptive hierarchical video multicast solution in [27] considered jointly application, data link,
and physical layers, where channel dependent Auto Rate
Selection was proposed. To combat packet losses in multicast, a layered hybrid Automatic Repeat reQuest scheme
was proposed in [28], where operating point for the multicast group was selected by a Nash bargaining game. The
approach in [29] for video unicast/multicast over wireless
proposed to minimize the resource usage while satisfying the diverse QoS requirements. The adaptive multicast
in [30] maintains the highest sustainable transmission rate
with suitable forward error correction (FEC) to maximize
the received video quality. These approaches however did
not address channel dependent SVC rate adaptation, MCS,
and receiver constraints.
The approach in [9] proposed to enable the heterogeneous receivers render the appropriate sub-streams by time
slicing technique in DVB-H for energy saving. This study
derived the rate allocation to different layers from uniform, linear, or exponential distribution. But in actuality
the rate of the layers depends on the encoding parameters
(e.g. frame rate, quantization level, and spatial resolution).
Also, the quality of received video and the effect of channel
condition were not studied here.
A recent study [15] considered heterogeneous broadcast
users, where an objective (temporal-spatial rate) distortion
metric was used based on Principal Component Analysis
distance between frames, and optimal layer broadcasting
policy was obtained to maximize the utility. However, it did
1524
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 7, JULY 2014
not consider channel adaptive scalability of SVC content,
dynamic physical resource allocation, and energy saving at
the receiver.
Adaptive modulation and coding (AMC) has been
widely employed to effectively combat the channel dynamics and maximize physical layer data rate. In the context of
video broadcast over wireless there are a few recent works
(e.g., [12], [10], [13], [14]) which have used AMC in different
forms and with different objectives.
The AMC approach in DVB-H applications in [12], which
we call simple MCS scheme, decides on adaptation based on
the broadcast receiver with an acceptable weakest signal
strength and uses the same MCS for all SVC layers. In the
AMC approach for DVB-H transmission [10], which we call
fixed MCS scheme, different layers are assigned a predetermined fixed MCS. This scheme results in saving of power at
both data reception and processing. However, in this work,
the adaptation is merely on the basis of transmitted frame
arrangement which is organized in terms of weaker and
incremental codes; it does not incorporate video encoded
data rates or the use of SVC to support heterogeneous
users.
Unlike in DVB-H, transmission rate optimization in LTE
MBMS is not based on time slicing. The adaptive MCS
in [13] is in context of orthogonal frequency-division multiple access (OFDMA). The approach in [14] is also for LTE
and WiMAX systems, where cooperative reception from
multiple BSs is utilized following the Single Frequency
Network principle. In all these AMC approaches, device
limitations were not considered, thus the application layer
encoding rate, and hence MCS is not affected by the
heterogeneity of users in the network.
3
S YSTEM M ODEL
3.1 Overview of the System
A single-cell broadcast scenario is considered. Multimedia
content delivery is done from the BS and managed jointly
with a connected media server. The wireless user equipments (UEs) have varying display resolution and battery
capabilities. Based on the users characteristics in the cell
and their SNRs, the media server suitably encodes the
source content in H.264/SVC standard of DVB-H. The
broadcast over the physical channel is OFDM-based. A UE,
depending on its current status, may choose to receive all or
part of the broadcast content (layers) by exploiting the timesliced transmission feature of DVB-H. Fig. 1 illustrates a
representative system, where L layers and T user types are
considered. For example, L = 14 in the standard ‘Harbor’
video sequence.
Definition 1. UE type τ is characterized by the spatial resolution R of a UE display and battery power, which are
device-specific. In a system with T types (1 ≤ τ ≤ T ) of
UE, Ri > Rj if j < i. For example in a system with 3 types
of users (i.e., T = 3), R3 > R2 > R1 .
SVC supports three types of scalability: spatial, temporal,
and SNR-based. Spatial scalability is governed by display
resolution of the UE (e.g., QCIF, CIF, D1), temporal scalability
is related to the frame transmission rate (e.g., 1.875 fps to 30
fps), and SNR scalability is linked with the SVC coding rate
Fig. 1. Example of the DVB-H system, where L-layer SVC content
and T types of UEs are considered.
as a function of the SNR experienced by the various UEs. A
detailed overview of H.264/AVC scalable video extension
is given in [18].
In our study, the supported spatial resolutions considered are QCIF (quarter common intermediate format, with
display resolution 176 × 144 pixels), CIF (common intermediate format, resolution 352 × 288 pixels), and D1 (D-1
digital recording video standard, resolution 704×576 pixels)
formats, which serve three types of users (i.e., T = 3). Apart
from spatial resolution of the individual video frames,
variable frame rate is also considered for the transmitted
video.
Definition 2. Layer l (1 ≤ l ≤ L) of a SVC content with a total
of L layers implicitly has its priority Pl in an inverse order
with respect to the other layers, i.e., Pi > Pj , if i < j.
If L = 14, following definition 2, P1 > P2 > · · · > P14 . If
a type i UE finds useful to display content up to the layer
l(i) ( ≤ L), then l(i) < l(j) for i < j.
SVC encoding generates different layers: base layer
(layer 1) and enhancement layers. Layer 1 is the most
important that needs to be received by all the UEs for the
basic minimum quality. The other layers when received by
a UE improve the reception quality by increasing the frame
rate and/or resolution at the playback stage.
3.2 Proposed DVB-H System Framework
The proposed overall system architecture is illustrated in
Fig. 2. The server encapsulates the SVC encoded data in
real-time transport protocol (RTP) format to IP packets
and sends them to the BS. The BS comprises of the IP
encapsulator, DVB-H modulator, and the radio transmitter. IP encapsulator puts the IP packets into multiprotocol
encapsulation (MPE) frames and forms MPE-FEC for burst
transmission as per the time slicing scheme (Section 4.2).
The DVB-H modulator employs an adaptive MCS selection
(Section 4.6) for the layered video content and sends it to
the radio transmitter for broadcast.
The SVC encoding and MPE-FEC framing operations are inter-dependent and jointly optimized based on
some underlying parameters (user, channel, and layer
SINGHAL ET AL.: JOINT OPTIMIZATION OF USER-EXPERIENCE AND ENERGY-EFFICIENCY IN WIRELESS MULTIMEDIA BROADCAST
1525
Fig. 3. Signal quality-based grouping example, with 3 UE types.
Fig. 2. Proposed DVB-H system architecture components.
information). The optimized video encoding parameters are
obtained through a game theoretic approach and stored in
a central database. The UE and channel aware user grouping
is discussed in Section 4.1, and SVC parameter optimization
game is detailed in Section 4.5.
The UE informs its capabilities while subscribing to the
broadcast service and also time-to-time updates its signal
strength to the BS. It also has a power manager that helps to
take advantage of the time slicing scheme and save energy
based on its remaining power.
Definition 3. A user class c (1 ≤ c ≤ C) defines the capability
of receiving the number of layers which is dictated by channel
rate constraint experienced by the UE at a given instant of
time. C is the total number of user classes.
If a UE can receive up to ls useful layers, it belongs to
class c = ls . Thus, a user class is dynamically associated to a
UE and is upper bounded by its resolution. If the number
of useful layers of a type i UE with resolution Ri is l(i) ,
(l(i) ≤ L), then it can be in class c such that 1 ≤ c ≤ l(i) . For
a UE type k with the highest resolution Rk , l(k) = L. In that
case, L = C and 1 ≤ c ≤ L.
The parameters updated by the BS in the database are: C,
the number of user classes; Nc , the number of users in class
c (1 ≤ c ≤ C); and R, the OFDM channel rate (expressed
in bps). The parameters updated by the video server in the
database are: L, the number of layers in the encoded SVC
content; rl , the rate of layer l (1 ≤ l ≤ L); and b, the burst
size of the base layer (measured in bis).
The proposed system performance optimization
involves: i) grouping of users, ii) game theoretic formulation to obtain SVC encoding parameters, iii) time slicing
at data-link level transmission, and iv) adaptive MCS
allocation to the SVC layers. These are discussed next.
4
P ERFORMANCE M ODELING AND
O PTIMIZATION OF THE P ROPOSED S YSTEM
4.1 Grouping of Users
User grouping is based on the respective UE resolution
capabilities and received SNR. A UE capability is determined by the BS at the time of service subscription, when
Fig. 4. Time slicing based DVB-H broadcast scheme.
the UE sends its type information, i.e., the number of layers
it wants to receive. The UE periodically updated its channel
condition to the BS through the uplink channel.
(τ )
Definition 4. User group glz refers to the UEs of type τ (1 ≤
τ ≤ T ) in zone z(1 ≤ z ≤ Z ) that have requested for l(τ )
layers. Z is the number of concentric zones around a BS.
The coverage region of a BS is comprised of concentric
zones, as shown in Fig. 3, with the SNR thresholds defining
zone boundaries. For a SVC content with L layers, 1 ≤ l(τ ) ≤
L, with QCIF resolution l(τ ) = 4, and those with CIF and D1
resolutions are respectively 9 and 14. In the user-grouping
example of Fig. 3, three UE types and three zones (based on
three supported MCS levels in DVB-H [31]) are considered.
The groups in this example are: g4z1 = {U2, U4, U6}, g9z1 =
4
9
14
{U1, U5}, g14
z1 = {U3}, gz2 = {U8, U11}, gz2 = {U7, U9}, gz2 =
4
9
14
{U10}, gz3 = {U13, U15} , gz3 = {U14} and gz3 = {U12, U16}.
4.2 Time Slicing as an Energy Saving Measure
Time slicing approach allows discontinuous reception at the
UEs, thereby facilitating the UE to turn-off the radio when
not receiving data bursts and hence saving energy.
Definition 5. Energy saving (ES) is calculated as the ratio of
the time duration for which the UE’s radio components are
turned-off over the total time of the video transmission cycle.
The broadcast channel rate is considered R (bps). The multimedia content is encoded into L layers. For decoding the
layer l (1 ≤ l ≤ L) the UE first needs to correctly receive
and decode all layers `l, 1 ≤ `l < l. Video layer l is allocated
rate rl (bps), such that Ll=1 rl ≤ R.
In time slicing-based layered broadcast, the UEs know
apriori the specific layer constituted in a MPE-FEC frame
(burst). As shown in Fig. 4, each layer corresponds to a
different burst within the recurring window. This allows a
UE to safely skip the bursts containing the layers that are
irrelevant to it, and thereby save energy. Each MPE-FEC
1526
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 7, JULY 2014
frame consists of two parts: Application Data Table that
carries the IP packet, and an R-S (Reed-Solomon coding)
Data Table that carries the parity bits.
Given a channel rate R and base layer burst size b bits,
the burst size of layer l is proportionally set to b · rl /r1 bits.
The recurring window size is the total burst size of all the
layers, given as: Ll=1 b · rl /r1 = b·R
r1 bits. Hence with respect
to starting time ofthe base layer burst, the start time of the
b·
l−1
i=1 ri /r1
layer l burst is:
sec.
R
If a user is currently in class c, the energy saving factor
of that user at that time instant would be:
c
ri
H · c · r1
(1)
ESc = 1 − i=1 −
,
R
b
where, in general 1 ≤ c ≤ l(τ ) , for a type τ UE, H is the
overhead duration (typically 100 ms [9]).
4.3 Video Quality Model
The video quality Q(q, t) is a parametric function that best
approximates the Mean Opinion Score (MOS). MOS is a
subjective measure that indicates the user QoE level. MOS
5 refers to ’excellent’ quality, 4 is ’good’, 3 is fair, 2 is ’poor’,
and 1 is ’bad’. The parameters for the quality model are
specific to a video based on its inherent features. The quality parametric model in [32] is specified with video specific
parameters λ and g. For a given spatial resolution, Q(q, t)
is a function of the quantization parameter QP and frame
rate t, as follows:
Q(q, t) = Qmax · Qtc (t) · Qq (q), with
Qtc (t) =
(2)
1 − e(−λ·t/tmax )
,
1 − e−λ
e(−g·q/qmin )
, and q = 2(QP−4)/6 .
e−g
Qmax is the maximum quality of the received video at the
UE when it is encoded at minimum quantization level
qmin and at the highest frame rate tmax . To normalize, we
consider Qmax to be 100%.
To comprehensively study the video quality in the
proposed system framework, we consider three representative video sequences: ‘Harbor’, ‘Town’, ‘Tree’, which
cover a wide spatial and temporal perceptual information
space [33]. In particular, the ‘Harbor’ video represents a
sequence with sharp edges (high spatial variations) but
having a relatively slow motion (low temporal variations),
‘Town’ has high spatial and temporal variations, whereas
‘Tree’ has low spatial and temporal variations in first half
and high spatial and high temporal changes in the later half.
Fig. 5 captures the effect on quality Q of the three different
video sequences at different QP. The trends of variation of
Q (which represents QoE) are observed to be quite similar
in all these video sequences. Also, the plots indicate that
the quality is a concave function of QP.
Qq (q) =
4.4 Energy Saving Versus Quality Trade-Off
As noted in Section 4.2, the energy saving is a function of
rate allocation to the layers. We now consider the scalability
factors as parameters in the rate allocation at the source
encoding stage. The parametric rate model, as in [34],[35],
Fig. 5. Quality versus quantization parameter at different frame rates
with the three standard video sequences, namely, ‘Harbor’, ‘Town’,
and ‘Tree’.
again is a function of the quantization level q, frame rate
t, and spatial resolution s. The parameters θ , a and d, here
are video specific.
Rc (q, t, s) = Rmax · Rtc (t) · Rq (q) · Rs (s), with
(3)
1 − e(−θ ·t/tmax )
, Rq (q) = ea·(1−q)/qmin ,
s d
and Rs (s) =
, d < 1.
smax
Rtc (t) =
1 − e−θ
Here, Rmax is the maximum bit rate of the video sequence
with minimum quantization level qmin , maximum frame
rate tmax , and maximum spatial resolution smax . By using
these rate parametric model equations for energy saving
analysis (i.e., in (1)), the energy saving for class c users of
type τ (1 ≤ c ≤ l(τ ) ) is given as:
ESc = 1 −
Rc (q, t, s) H · c · R1 (q, tmin , smin )
−
.
R
b
(4)
To study the impacts of SVC quantization parameter and
time slicing scheme on the energy saving, we again consider the three representative video sequences: ‘Harbor’,
‘Town’, and ‘Tree’. Fig. 6(a) shows the variation of normalized average energy saving with the change in quantization
parameter QP. It is notable that, in all the three cases, an
increase in QP results in a higher energy saving. Also a
decrease in t (by DRX mechanism) and smaller spatial resolution of the video sequence results in more energy saving
for the UEs. It is also observed that the nature of the energy
saving is concave with respect to QP. For the three video
sequences, Fig. 6(b) shows the normalized energy saving
with respect to the layers received by the UEs. Also, a
higher QP (i.e., higher q and hence a lower allocated rate)
corresponds to a higher energy saving. Hence, there is a
clearly evident trade-off between the energy saving and
quality for a specific value of quantization level q. Also
different type of users have different energy savings and
QoE requirements.
It is observed from Figs. 5 and 6 that, the quality and
energy saving performances of the three representative test
sequences (‘Harbor’, ‘Tree’, ‘Town’) follow similar trends
with respect to the variations of quantization parameter,
frame rate, spatial resolution, and the number of layers
transmitted. Thus, the proposed framework in the paper
and the optimizations (discussed subsequently in this section) should generically hold true for any possible SVC
video sequences. Therefore, our remaining performance
results are discussed with respect to only one representative
test sequence (‘Harbor’ video).
SINGHAL ET AL.: JOINT OPTIMIZATION OF USER-EXPERIENCE AND ENERGY-EFFICIENCY IN WIRELESS MULTIMEDIA BROADCAST
1527
Fig. 7. Examples of utility plots of individual class of users for the
standard ‘Harbor’ video sequence.
Fig. 6. Energy saving performance at different SVC quantizations
and time slicing schemes for the three video sequences: ‘Harbor’,
‘Town’, and ‘Tree’. (a) Effect of quantization parameter QP. (b) Effect
of number of layers transmitted.
4.5 Energy Saving and Quality Optimization Game
Based on the energy saving and quality trade-off that
depends on the quantization level q, we now formulate
a cooperative game to obtain the optimal video encoding parameters. Note that, the development in Section 4.4
demonstrates the possibility of an optimal SVC encoding
from the individual user’s perspective. However it does not
provide an insight to the encoding optimality for broadcast
when there are different user class in different proportions.
Here, we address this optimization aspect.
This optimization game jointly accounts for the users
of different classes (Definition 3) as well as the fraction of
users in each class. The game is defined below:
Players: Class c comprising of a set of users who can be
served up to l = c layers, where 1 ≤ c ≤ l(τ ) , τ = 1, 2, 3.
(Recall that, c is dynamic, computed at the BS, depending
on the UE type τ and their individual SNRs.)
Strategy: Quantization level q used by the SVC encoder
for encoding the source video. Optimum q determines the
rate distribution (i.e. the minimum bit rate) rl for the different layers l of the SVC content, that satisfy the users’ ES
and quality requirements.
Utility: For class c the utility is defined as: uc =
(ESc (q, t))αc · (Qc (q, t))βc , where αc , βc are the parameters
for a particular class of users based on their emphasis on
energy saving or quality, with αc +βc = 1. The higher the αc
value is, the higher is the emphasis on energy saving by the
users in that class. On the other hand, the higher the value
of βc is, the more will be the emphasis on receiving higher
quality video. Here, for class c, energy saving ESc (q, t) is
given in (4) and the quality value Qc (q, t) is given in (2).
We use multiplicative exponent weighting (MEW) in
defining the utility uc instead of simple additive weighting
(SAW), because in SAW based optimality poor value of
a parameter can be outweighed by a very good value
of another parameter. Instead, MEW penalizes alternatives
with poor parameter values more heavily. For example, if
energy utility is near zero (which means the UE consumes
a lot of energy), the MEW based utility function avoids
selecting this because it is multiplicative, whereas the SAW
utility may end up choosing this case of near-zero energy
but very high quality.
In Fig. 7, some examples of utility function with QP variation are shown. Quantization level q is can be obtained
from QP (see (2)). The (αc , βc ) combination shown are the
optimum values that achieve the maximum possible utility for the individual user class c. The plots indicate that
for each class, the considered utility function is concave in
nature in terms of the QP.
Since the number of users in a class impacts the overall
system utility, we define a modified utility function. If there
are Nc users in class c, the modified utility is:
u`c = Nc (ESc (q, t))αc · (Qc (q, t))βc .
(5)
The objective is to maximize the total average utility for the
system,
⎛
⎞
l(τ )
T u` ⎟
⎜
⎜ τ =1 c=1 c ⎟
⎟.
Utotal = max ⎜
(6)
⎜ T l(τ )
⎟
⎝ ⎠
Nc
τ =1 c=1
Before we proceed further, we prove the concavity of the
utility functions in (5) and (6).
Proposition 1. The utility function of a class c, defined in (5),
and the system utility, defined in (6), are strictly concave
functions of QP in the range [QPmin , QPmax ].
Proof. Given two functions f1 (x) and f2 (x), a function
φ(x) = f1 (x) · f2 (x) is said to be strictly concave and
has a unique maxima in [xmin , xmax ] if the following
conditions hold [36]:
(1) f1 (x) < 0 and f2 (x) < 0, i.e., f1 (x) and f2 (x) are
concave functions of x ∈ [xmin , xmax ], (2) f1 (x), f2 (x) are
non-negative, and (3) f1 (x) · f2 (x) < 0 in [xmin , xmax ].
1528
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 7, JULY 2014
Since φ (x) = f1 (x) · f2 (x) + 2 · f1 (x) · f2 (x) + f1 (x) · f2 (x),
by the above conditions φ (x) is negative in [xmin , xmax ].
Hence it is concave down with a maxima at
k ∈ [xmin , xmax ], s.t.φ (k) = 0.
In our context the two functions are: ESc (q, t)αc and
Qc (q, t)βc . Since, the proof is generic and holds true
∀ c ∈ [1, l(τ ) ], and t, s are constant values for any
class c, the variable over which the optimization is
carried out is QP, which is related to q by (2). Thus
the two functions can be written as f1 (x) = ESc (QP)αc
and f2 (x) = Qc (QP)βc , and the interval of concavity is
[QPmin , QPmax ]. We want to show the concavity of the
utility function and joint optimization of the system in
terms the best suited QP for the video encoding.
Firstly, we prove Qc (QP)βc is concave in
[QPmin , QPmax ]. The proof is as follows:
=
(Qq (q) · Qmax · Qtc )βc , where
Qc (QP)βc
QP ∈ [QPmin , QPmax ].
Then, from (2):
(QP−4)/6 )/2(QPmin −4)/6
e(−g·2
Qc (QP)βc =
e−g
·D
βc
, where D is
a constant with respect to the variable QP and is given
as D = Qmax · Qtc (t).
For obtaining the derivative, denote QP = x,
=
xmin , QPmax
=
xmax . Also let
QPmin
w = 2((x−4)/6) /2((xmin −4)/6) . Then we have:
dQc (w)βc
dw
dQc (x)βc
dQc (w)βc dw
=
·
, where
(7)
dx
dw
dx
−g·w
βc
e
Qc (w)βc =
·D
,
e−g
dw
w
= (−βc · g) · e−βc ·g·(w−1) · D, and
= .
dx
6
It is evident from (7) that, for xmin = QPmin = 1, xmax =
QPmax = 51:
dQc (x)βc
< 0, ∀ x ∈ [xmin , xmax ]
dx
(8)
Differentiation (7) again with respect to x we have,
d2 Qc (x)βc
1
× (βc · g) · e−βc ·g·(w−1)
=
β
·
g
−
c
6
dx2
·D · 2
Since βc · g <
(9),
d2 Qc (x)βc
dx2
<
1
6
((x−4)/6)
/(6 · 2
((xmin −4)/6)
(9)
)
(βc (max) = 1, g = 0.06), from
0. Thus Qc (QP)βc is concave in
[QPmin , QPmax ].
We now prove that ES(QP)αc is concave in
[QPmin , QPmax ].
αc
From (4), ESc (QP)αc = 1 − Rc (QP)
− H·c·Rb1 (QP) ,
R
where QP ∈ [QPmin , QPmax ].
=
it implies that
ESc (QP)αc
αc
From (3),
(QPmin −4)/6
(QP−4)/6
a(1−2
)/(2
)
) · P , where P is a con(1 − e
stant with
QP. Using (3), P is obtained
respect to variable
Rmax ·Rtc (t)·Rsc (s)
H·Rmax ·Rtc (tmin )·Rs (smin )
−
.
as: P =
R
b
Again, for the derivative we denote QP = x, QPmin =
xmin , QPmax = xmax . Let v = ea(1−2
(x−4)/6) )/2(xmin −4)/6
· P.
Fig. 8. Total average system utility with the ‘Harbor’ sequence.
By simplifying we have,
dESc (x)αc
dESc (v)αc dv
=
· , where
dx
dv
dx
αc
ESc (v) = (1 − v)αc and
(x−4)/6
(x
)/2 min
dv
aea(1−2
=−
dx
(2(xmin −4)/6 )
−4)/6
Since v < 1 ∀ x ∈ [xmin , xmax ],
This implies that,
·P·
dv
dx
(10)
2(x−4)/6
· ln 2.
6
< 0 and
dESc (v)αc
dv
< 0.
dESc (x)αc
> 0, ∀ x ∈ [xmin , xmax ].
(11)
dx
On similar lines as in (7)-(9), it is observed that
d2
(ESc (x)αc ) < 0, ∀ x ∈ [xmin , xmax ]. This implies that
dx2
ES(QP)αc is concave in [QPmin , QPmax ].
Thus, Condition (1) is shown to be true. Condition
(2) holds in the proposed scheme as per the basic
design of the system, since ESc (QP)αc and Qc (QP)βc
are always positive. From (8) and (11), the product
ESc (QP)αc · Qc (QP)βc is negative. So Condition (3) also
holds true. Hence, the utility of class c, uc is proven to
be strictly concave and has a maxima in the given range.
Note that, although the exact value of the maxima for a
class of utility function and the corresponding value of
QP can be easily obtained from the above development,
these are not of our interest here.
As shown in [37], [38], a non-negative linear combination of strictly concave functions is also strictly concave.
Since the system utility in (6) is a non-negative linear combination of the utilities of all the individual
classes that are already proven to be strictly concave,
the system utility in (6) is a strictly concave in the
given range. This implies the existence of a unique solution that maximizes the utility in the joint optimization
formulation.
In terms of algorithmic complexity, the joint energysaving and video quality optimization has a complexity
of O(lT ), where lT is the number of SVC layers being
broadcast for the highest resolution type T .
Fig. 8 shows that, the sum total weighted average utility
for all classes is a concave function of QP with a unique
maxima. This plot corresponds to 60% type 1 users, 30%
type 2 users, and 10% type 3 users in the system, with their
corresponding random location dependent SNR feedbacks
accounted at the BS to determine the user classes c and the
corresponding Nc values. (Different traffic scenarios and the
SNR-channel rate relationships are given in Tables 1 and 2,
respectively.) The maxima of this scenario corresponds to
QP = 30.
SINGHAL ET AL.: JOINT OPTIMIZATION OF USER-EXPERIENCE AND ENERGY-EFFICIENCY IN WIRELESS MULTIMEDIA BROADCAST
TABLE 1
Simulation Scenarios, with Variable Ratios of user Type
4.6 Adaptive Modulation and Coding Scheme
As noted in Section 3, in our approach, besides user-andchannel aware SVC rate optimization at the application
layer and time slicing at the link layer, at the physical layer
adaptive MCS is applied which is optimized for enhanced
energy efficiency and network capacity. Clearly, this adaptation is a function of the heterogeneous users composition
in a cell and the dynamic physical channel rate constraint.
Physical channel dynamics is accounted in a slow (shadow
fading) scale to avoid high bandwidth overhead of frequent
channel state feedback and computation of coding and MCS
optimizations at the BS as well as the video server.
The total number of users in the cell that are subscribed
to the broadcast service is taken to be N. The different MCSs
are considered to be m = 1, . . . , M (for example, m = 1
represents QPSK with code rate = 1/2, and so on). The
SVC encoded video is considered to have L layers. In our
formulation, Rm represents the data rate provided by MCS
m, rl is the rate allocated for layer l (l = 1, . . . , L). If a layer l,
can be served by the BS to the users, then we set lserved = 1,
else it is set to 0. We have used an indicator function χl,m
that takes a value 1 if layer l is modulated with MCS m and
takes a value 0 otherwise. The value of ml , l = (1, . . . , L)
specifies the MCS used for layer l subject to lserved = 1. For a
user to be able to decode any layer, it is necessary to have
received all the layers lower than the current layer. Only
then the layer is said to be valid for the user. The number
of valid layers for any user is denoted by:
M
lsj = max{l | ∀ `l ≤ l ≤ l(τj ) ,
j
`
m=1
χ`l,m` = 1}.
lsj is the maximum number of continuous layers modulated with 1 to Mj starting from base layer, where Mj is the
highest possible modulation level that jth user can receive,
such that these layers are either equal to or less than the
requested number of layers by the user (l(τj ) ) based on its
type τj , 1 ≤ τj ≤ T . The received rate for user j is given
l
by: rj =
sj
l=1
rl .
The utility for the user j is defined as a general function
of its received rates, requested quality rates and maximum possible feasible received rate based on its channel
conditions, i.e. SNR with shadowing at any given time:
(τj )
(τj )
Uj (rj , rl , rj(SNR) ). Here, rl
1529
corresponds to the rate
requested by the user for its desired maximum quality level.
So it is based on the maximum number of layers l(τj ) of the
SVC content requested by the user of type τj . rj(SNR) corresponds to the maximum rate that the user would be able
to receive if the user alone was present in the network and
optimization of the MCS was to be just based on this user’s
channel conditions (i.e., its experienced SNR).
The user j’s utility is defined as:
(τj )
l
, rj(SNR) )
Uj (rj , r
(τj )
l
= Q(rj ) − Q(min{r
, rj(SNR) }),
(12)
where Q(rj ) is the quality value based on the parametric
model given in (2). Since the possible rates that a user j
can receive is from a set of possible layer rates, if a user is
able to receive lsj layers, such that rj ≥ rls , with layer lsj
j
having a frame rate tls and a quantization level qls , then,
j
j
Q(rj ) = Q(qls , tls ).
j
j
The objective is to maximize the total system utility,
(τj )
i.e., max{ Uj (rj , rl , rj(SNR) )} subject to (5), (6), and the
j∈N
following constraints:
rj ≥ r1 , given that, rj(SNR) ≥ r1
M
χl,m ≤ 1,
∀j ∈ N
l = 1, . . . , L
(13)
m=1
The first set of constraints mentioned in (13) states that
for every user j ∈ N the rate received should be at least
greater than or equal to the rate of the base layer r1 , for
the condition that rj(SNR) > r1 , i.e. the channel condition of
the user j supports a rate greater than that required for the
base layer. The second set of constraint in (13) uses integer relaxation, which states that for a given video layer l,
it can be modulated and coded with at most one MCS. It
is important to note that a layer l may not even be modM
ulated with any MCS, i.e.,
χl,m = 0. In such a case the
m=1
layer l is not transmitted. The users experiencing extremely
bad channel conditions with rj(SNR) < r1 will not be able
to receive any layer, since they are experiencing the SNR
below the minimum SNR threshold for the most basic MCS
(e.g., QPSK with code rate of 1/2).
The different supported MCS have a minimum SNR
threshold (γm , for MCS m = 1, . . . , M) under the given DVBH standard specifications, based on the quasi error free
reception and MPE-FEC error rate of 5% with a BER value
of 10−12 [31]. The rates corresponding to SNR threshold of
TABLE 2
MCS Parameters with Gaussian Channel Model and Guard Interval GI = 1/4 in DVB-H Standard [31]
1530
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 7, JULY 2014
Algorithm 1 Adaptive MCS selection for SVC layers
(τj )
Input: L, γm , Rm , rl , rj(SNR) , rl , ∀ m = 1, . . . , M, l =
1, . . . , L, and j = 1, . . . , N
Pseudocode:
1) Initialize variables: Utotal = 0 and χl,m = 0, ∀ l =
1, . . . , L, ∀ m = 1, . . . , M
2) for each l = 1 to L
if RM < rl then Set lserved = 0
else
for each i = 1 to M
if rl < Ri and ml−1 ≤ i then
Set ml = i, χl,i = 1, and lserved = 1
go to 3
3) for each user j = 1 to N
Using ml (l = 1, . . . , l(τj ) ) and lserved , find rj
Fig. 9. Mapping between MOS and parametric video quality.
(τj )
Compute Uj (rj , rl , rj(SNR) ) using (12)
(τj )
l , r
Utotal = Utotal + Uj (rj , r
)
j(SNR)
Output: χl,m , ml , lserved , ∀ l = 1, . . . , L and m = 1, . . . , M.
Fig. 10. Spatial-temporal scalable layer structure used in system
different MCS are given by:
1 ≤ m ≤ M,
` > m, 1 < m
` ≤M
γm` > γm , where MCS m
` ∈ [1, M], m
` > m.
Hence, Rm` > Rm ∀ m, m
simulations.
Rm = B. log2 (1 + γm ) ,
(14)
The proposed MCS assignment algorithm is summarized
in Algorithm 1.
The adaptive MCS algorithm has O(L · M + N) complexity, where L is the number of SVC layers broadcast,
M is the number of MCS levels, and N is the total number of users. The proposed approach ensures that, with
optimal MCS allocation for all the SVC layers, Q(rj ) ≥
(τj )
l , r
Q(min{r
}), ∀ j ∈ [1, N], leading to maximum
j(SNR)
system utility Utotal .
4.7 Video Reception Quality Measure
For a fair comparison of the quality of reception performance of the different competitive strategies, we define a
video reception quality measure.
Definition 6. In a system having T types of users, with the
highest number of layers l(τ ) that a type τ user (1 ≤ τ ≤ T )
is capable of receiving
and the corresponding reception quality
denoted by Q l(τ ) , the weighted average video reception
quality, or the Q measure is expressed as:
Q=1−
l(τ ) T (τ )
Q l(τ ) − Q (ls ) Nls (15)
1
l(τ )
T τ =1ls =0
rate t (Q (ls ) = Qls (q, t), i.e., the quality value corresponding
to layer ls (1 ≤ ls ≤ l(τ ) ) of SVC content).
With respect to a given UE type, Q is indicative of the
difference in actually experienced video reception quality
with respect to its highest reception capability. It is a performance metric that indicates the QoE of the broadcast
users in a given heterogeneous users distribution. Being a
weighted average, it also indicates the proportion of total
number of users that are served with a specified video quality level. Hence, a higher value of Q measure signifies that a
higher proportion of total number of users are being served
in the cell with a higher video reception quality. In our system example, T = 3, and l(τ ) = 4, 9, 14 respectively, for
τ = 1, 2, 3.
Further, it may be noted that the parametric quality
measure Q(q, t)and hence the weighted average quality
measure Q, that we use to characterize the transmission
strategies, have a direct relationship with the subjective
measure MOS [32], given as: MOS = 4 × Q(q, t) + 1.
Thus, numerically, Q(q, t) = 0 corresponds to MOS = 1,
Q(q, t) = (0.0 − 0.25] corresponds to MOS = 2, Q(q, t) =
(0.25 − 0.5] corresponds to MOS = 3, Q(q, t) = (0.5 − 0.75]
corresponds to MOS = 4, and Q(q, t) = (0.75 − 1.0] corresponds to MOS = 5. This mapping between QoE measure
of video quality (MOS) and parametric video quality Q(q, t)
is shown in Fig. 9.
(τ ) τ =1 ls =0
Nls
(τ )
where Nls is the number of type τ users actually receiving
ls number of layers, with the corresponding quality measure
Q (ls ). Q (ls ) = 0 if ls = 0, i.e. when no layers are received.
Q(l(τ ) ) and Q(ls ) are obtained based on the parametric
model in (2), as a function of quantization level q and frame
5
S IMULATION S ETTINGS
For the simulation purpose and in order to encode the SVC
streams, we have used the SVC encoder reference software
JSVM_9_19_12 [39]. In the considered scenario, scalable
video covers three levels of spatial resolution formats: QCIF,
CIF, and D1, serving three type of users, and five temporal resolutions: 1.875, 3.75, 7.5, 15, 30 fps (see Fig. 10),
SINGHAL ET AL.: JOINT OPTIMIZATION OF USER-EXPERIENCE AND ENERGY-EFFICIENCY IN WIRELESS MULTIMEDIA BROADCAST
1531
TABLE 3
Simulation Parameters
Fig. 11. Normalized energy saving versus reception quality of the
different types of users.
which serve the users in variable channel conditions. The
sample ‘Harbor’ video sequence with 300 frames was taken
for evaluating the proposed framework. For this video, the
parameters λ, g, θ , a, and d in (2)-(3) are respectively found
to be 7.38, 0.06, 1.429, 1.551, and 0.845. Fig. 10 shows the
flexible layer structure with each coordinate representing
different spatial and temporal resolutions. Note that there
are two possible layer routes for the hierarchical broadcast
reception.
We consider a single-cell video broadcast network with
500 randomly distributed users belonging to three user
types. Six simulation scenarios are considered with different ratios of user type distributions as listed in Table 1. The
users belonging to type 1 require QCIF format video, those
belonging to type 2 require CIF format, and the ones of
type 3 need D1 format.
The first step is to obtain the optimized SVC encoding
parameters as a function of the user types distribution in
the system using the proposed cooperative game. The outcome of the game is the optimized adaptive video coded
sequence with the optimal rate allocation for each layer,
such that it aids the energy saving for type 1 users and the
quality for type 3 users.
The rate allocation is followed by the adaptive MCS
allocation for the different SVC layers that are transmitted
in a time-sliced arrangement. Note that, since the optimal
rate allocation to different SVC layers rl , 1 ≤ l ≤ L is
a function of the user type distribution ratios, the timeslice allocation in (1) is also accordingly a function of user
type distribution ratios. The adaptive MCS in our approach
is additionally governed by the SNR experienced by the
various user groups.
As per the DVB-H specifications [23], the minimum
SNR threshold for each MCS for a given wireless channel and the corresponding channel rates with the relevant guard interval (GI = 1/4) are listed in Table 2.
The overall system simulation parameters considered are
listed in Table 3. The performance results are presented
below.
6
R ESULTS AND D ISCUSSIONS
6.1
Energy-Quality Trade-Off Performance With
Time Slicing Technique
Considering the different traffic scenarios as listed in
Table 1, we first analyze the energy–SVC quantizationdependent quality trade-off. Fig. 11, illustrates the energyquality (Q, given in (2)) trade-off for the three types of
users. It can be seen that, when the video encoding parameters are so chosen that the quality of the video at the UEs
is higher, the corresponding energy saving for the users
of all the three types is lower. However, under all the six
scenarios the energy saving for the type 1 users is the highest among the three type of users. This is primarily due
to the time slicing approach of transmission. Considerable
variation in the energy saving and quality values is evident when there is a remarkable change in proportion of
any particular type of user in the network. For instance,
in scenario 5 with 90% of type 1 users, the joint optimization approach results in energy saving of more than
90% for the UEs, with approximately 20% quality. This
is because, more than 90% users are energy-constrained
and the objective is to satisfy these users in terms of their
energy-saving.
It is also notable that, since each user has the independent control of time-sliced reception, even though the
high-end (e.g. type 3) users may not achieve the maximum
desired quality due to the system optimization for large
proportion of low-end (e.g. type 1) users, they can improve
the QoE by the time slicing flexibility.
6.2 Adaptive MCS Performance
We now study the MCS-dependent broadcast performance.
The proposed adaptive MCS is compared against the two
other schemes: simple MCS and fixed MCS.
Fig. 12(a) shows the average difference between the
user request for a certain number of layers and what they
actually receive under the three MCS schemes. Fig. 12(b)
shows that the total number of users receiving exactly
1532
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 7, JULY 2014
Fig. 14. Comparison of MCS allocation schemes in terms of service
capacity.
Fig. 12. MCS-dependent quality and capacity performance.
(a) Average difference between requested and received layers.
(b) Fully serviced users count under different MCS.
simple MCS scheme under the six scenarios with ‘ES+Q’
strategy are 12.6, 19.4, 14.1, 9.7, 7.2, and 24.4%, respectively. The average number of additional users served
with adaptive MCS under the six scenarios is 16.57% with
respect to the simple MCS and 16.63% with respect to the
fixed MCS.
6.3
Fig. 13. Effect of MCS and number of layers on the average
percentage of users served.
the requested quality is much higher with adaptive MCS
as compared to the two other schemes. The results are
averaged over several iterations with the number of users
varying between 400 and 500. It can be noticed that the
adaptive MCS outperforms the other two MCS schemes in
terms of the number of served users. Moreover by using
the adaptive MCS the received number of layers are very
close to the requested number of layers, reflecting a higher
amount of user satisfaction.
Fig. 13 captures the layer based values for the average
percentage of users being served in the broadcast network
for the six scenarios listed in Table 1 under the three MCS
allocation schemes. The results show that the adaptive MCS
allocation scheme outperforms the other schemes, by ensuring a higher percentage of users that are getting served
under all scenarios.
The composite gain achieved by the adaptive MCS under
the six different scenarios is illustrated in Fig. 14. The
ratio of the total number of users served by the proposed adaptive MCS is compared against the simple MCS
and fixed MCS. It can be noted that, among the three
schemes, the adaptive MCS ensures more number of users
served. In particular, the gain of adaptive MCS in terms
of number of additional users served over the fixed or
Energy-Quality Trade-Off With Optimized SVC
and Adaptive MCS
Fig. 15 presents a comparison of the three MCS along with
the three SVC optimization measures: joint energy saving
and quality (ES+Q), energy saving only (ES only), and quality only (Q only). The line plots indicate the number of
users served versus the number of SVC layers transmitted.
The adaptive MCS with ‘ES+Q’ is shown to perform better than the ‘Q only’ case. Although the ‘ES only’ serves a
higher number of users (see Table 4), the average reception
quality is very low (e.g., Q is 18.19% in scenario 1). This
means, in ‘ES only’ case a large proportion of users would
experience a low QoE. The uneven trend in the plots may
be due to random distribution of heterogeneous UEs in the
network.
The bar plots on the right in Fig. 15 capture the composite gain on number of users served in the three schemes
(‘ES+Q’, ‘ES only’, and ‘Q only’) with the three MCS strategies. ‘ES+Q’ with adaptive MCS serves a lesser number
of users in comparison with the ‘ES only’ case, but performs better with respect to ‘Q only’ case. The number of
users served in ‘Q only’ case is generally low (e.g., 35%
only in scenario 2). The results demonstrate that, with the
proposed adaptive MCS, the average reduction of number
of users served with ‘ES+Q’ is only 0.62% lower than that
in ‘ES only’ scenario. On the other hand, the number of
users served with ‘ES+Q’ is about 10.8% higher than that
with ‘Q only’ scenario.
Table 4 further quantifies the energy-quality trade-off
with the three optimization schemes: ‘ES+Q’, ‘ES only’, and
‘Q only’ where the weighted quality measure Q in (15) is
used. The table also includes the optimum quantization
parameters for the ES and Q trade-off game. Under the
six user-heterogeneity scenarios with adaptive MCS, when
compared with ‘ES only’ strategy, the ‘ES+Q’ strategy offers
on average, about 43% higher quality. The corresponding
trade-off on the amount of energy saving is only about 8%.
With respect to ‘Q only’ scenario, the ‘ES+Q’ scheme offers
about 17% extra energy saving as well as about 3.5% higher
quality performance.
SINGHAL ET AL.: JOINT OPTIMIZATION OF USER-EXPERIENCE AND ENERGY-EFFICIENCY IN WIRELESS MULTIMEDIA BROADCAST
1533
Fig. 15. Average percentage of users served with different MCS versus number of layers, and the aggregate performance.
TABLE 4
Average Energy Saving and Quality Performance with Adaptive MCS Under Different Traffic Scenarios
7
C ONCLUSION
This paper has introduced a novel cross-layer optimization solution to improve both the quality of user experience
(QoE) and energy efficiency of wireless multimedia broadcast receivers with varying display and energy constraints.
This joint optimization is achieved by grouping the users
based on their device capabilities and estimated channel
conditions experienced by them and broadcasting adaptive content to these groups. The optimization is a game
theoretic approach which performs energy saving versus
reception quality trade-off, and obtains optimum video
encoding rates of the different users. This optimization is a
function of the proportion of users in a cell with different
capabilities, which in turn determines the time slicing proportions for different video content layers for maximized
energy saving of low-end users, while maximizing the
quality of reception of the high-end users. The optimized
layered coding rate, coupled with the receiver groups’
SNRs, adaptation of the MCS for transmission of different layers, ensure higher number of users are served while
also improving users’ average reception quality. Thorough
testing has shown how the proposed optimization solution
supports better performance for multimedia broadcast over
wireless in comparison with the existing techniques.
1534
ACKNOWLEDGMENTS
This work has been supported by the Department
of Science and Technology (DST) under the Grant
SR/S3/EECE/0122/2010 and Indo-Ireland cooperative science program. The authors are thankful to the anonymous
reviewers for the insightful comments and valuable suggestions, which have significantly improved the quality of
presentation.
R EFERENCES
[1] Google, “The new multi-screen world: Understanding crossplatform consumer behavior,” Aug. 2012.
[2] Cisco, “Cisco visual networking index: Global mobile data traffic
forecast update, 2013–2018,” Feb. 2014.
[3] K. Rao, Z. S. Bojkovic, and D. A. Milovanoic, Multimedia
Communication Systems: Techniques, Standards, Networks, 1st ed.
Upper Saddle River, NJ, USA: Prentice Hall, 2002.
[4] S. Sesia, I. Toufik, and M. Baker, LTE UMTS Long Term Evolution,
from Theory to Practice. Chichester, U.K.: Wiley, 2009.
[5] G. Faria, J. Henriksson, E. Stare, and P. Talmola, “DVB-H: Digital
broadcast services to handheld devices,” Proc. IEEE, vol. 94, no. 1,
pp. 194–209, Jan. 2006.
[6] Digital Video Broadcasting (DVB); Framing Structure, Channel Coding
and Modulation for Digital Terrestrial Television, ETSI EN Std. 300
744, 2009.
[7] Digital Video Broadcasting (DVB); DVB-H Transmission System for
Handheld Terminals, ETSI EN Std. 302 304, 2011.
[8] Amendment 5: Enhancements for Higher Throughput, IEEE Std.
802.11n, Oct. 2009.
[9] C.-H. Hsu and M. M. Hefeeda, “Flexible broadcasting of scalable video streams to heterogeneous mobile devices,” IEEE Trans.
Mobile Comput., vol. 10, no. 3, pp. 406–418, Mar. 2011.
[10] T. E. Shabrawy and S. Wahed, “Adaptive modulation and coding for broadcast DVB-H systems,” in Proc. IEEE PIMRC, Tokyo,
Japan, Sep. 2009, pp. 1292–1296.
[11] O. Karimi and L. Jiangchuan, “Power efficient high quality multimedia multicast in LTE wireless networks,” in Proc. IEEE MASS,
Valencia, Spain, Oct. 2011.
[12] C. Atici and M. Sunay, “High data-rate video broadcasting over
3G wireless systems,” IEEE Trans. Broadcast., vol. 53, no. 1,
pp. 212–223, Mar. 2007.
[13] A. Alexiou, C. Bouras, V. Kokkinos, A. Papazois, and
G. Tsichritzis, “Modulation and coding scheme selection in multimedia broadcast over a single frequency network-enabled longterm evolution networks,” Int. J. Commun. Syst., vol. 25, no. 12,
pp. 1603–1619, Nov. 2012.
[14] P. Li, H. Zhang, B.-H. H. Zhao, and S. Rangarajan, “Scalable video
multicast with adaptive modulation and coding in broadband
wireless data systems,” IEEE/ACM Trans. Netw., vol. 20, no. 1,
pp. 57–68, Feb. 2012.
[15] W. Ji, Z. Li, and Y. Chen, “Joint source-channel coding and
optimization for layered video broadcasting to heterogeneous
devices,” IEEE Trans. Multimedia, vol. 14, no. 2, pp. 443–455, Apr.
2012.
[16] Y. Wang, J. Ostermann, and Y.-Q. Zhang, Video Processing and
Communications. Upper Saddle River, NJ, USA: Prentice Hall,
2001.
[17] Overview
of
the
MPEG-4
Standard,
ISO/IEC
Std.
JTC1/SC29/WG11. N4668, Mar. 2002.
[18] H. Schwarz, D. Marpe, and T. Wiegand, “Overview of the scalable
video coding extension of the H.264/AVC standard,” IEEE Trans.
Circuits Syst. Video Technol., vol. 17, no. 9, pp. 1103–1120, Sep. 2007.
[19] T. Wiegand, G. J. Sullivan, G. Bjntegaard, and A. Luthra,
“Overview of the H.264/AVC video coding standard,” IEEE
Trans. Circuits Syst. Video Technol., vol. 13, no. 7, pp. 560–576, Jul.
2003.
[20] T. Schierl, T. Stockhammer, and T. Wiegand, “Mobile video transmission using scalable video coding (SVC),” IEEE Trans. Circuits
Syst. Video Technol., vol. 17, no. 9, pp. 1204–1217, Sep. 2007.
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 13, NO. 7, JULY 2014
[21] G. Iwacz, A. Jajszczyk, and M. Czkowski, Multimedia Broadcasting
and Multicasting in Mobile Networks. Chichester, U.K.: Wiley, 2008.
[22] Scalable Video Coding (SVC), ISO/IEC Std. JTC1/SC29/WG11.
N9560, 2008.
[23] Digital Video Broadcasting (DVB); DVB-H Implementation Guidelines,
ETSI TR Std. 102 377 V1.3.1, 2009.
[24] M. Ghandi and M. Ghanbari, “Layered H.264 video transmission
with hierarchical QAM,” J. Vis. Commun. Image Represent., vol. 17,
no. 2, pp. 451–466, Apr. 2006.
[25] G. Xylomenos, “Group management for the multimedia broadcast/multicast service,” in Proc. IST Mobile Summit, Dresden,
Germany, Jun. 2005.
[26] Y.-C. Chen and Y.-R. Tsai, “Adaptive resource allocation for multiresolution multicast services with diversity in OFDM systems,”
in Proc. IEEE VTC, Barcelona, Spain, Apr. 2009, pp. 1–5.
[27] J. Villalon, P. Cuenca, L. Orozco-Barbosa, Y. Seok, and T. Turletti,
“Cross-layer architecture for adaptive video multicast streaming
over multirate wireless LANs,” IEEE J. Sel. Areas Commun., vol. 25,
no. 4, pp. 699–711, May 2007.
[28] Z. Liu, Z. Wu, P. Liu, H. Liu, and Y. Wang, “Layer bargaining:
Multicast layered video over wireless networks,” IEEE J. Sel. Areas
Commun., vol. 28, no. 3, pp. 445–455, Apr. 2010.
[29] Q. Du and X. Zhang, “Statistical QoS provisionings for wireless
unicast/multicast of multi-layer video streams,” IEEE J. Sel. Areas
Commun., vol. 28, no. 3, pp. 420–433, Apr. 2010.
[30] O. Alay, T. Korakis, Y. Wang, and S. Panwar, “Dynamic rate
and FEC adaptation for video multicast in multi-rate wireless
networks,” Mobile Netw. Appl., vol. 15, no. 3, pp. 425–434, Jun.
2010.
[31] J. T. J. Penttinen, P. Jolma, E. Aaltonen, and J. Väre, DVB-H Radio
Network. Chichester, U.K.: Wiley, 2009.
[32] Y. Wang, Z. Ma, and Y.-F. Qu, “Modeling rate and perceptual
quality of scalable video as functions of quantization and frame
rate and its application in scalable video adaptation,” in Proc. Int.
Packet Video Wksp., Washington, DC, USA, May 2009.
[33] Subjective Video Quality Assessment Methods for Multimedia
Applications, ITU-T Recommendations Std. P.910, Apr. 2008.
[34] S. Azad, W. Song, and D. Tjondronegoro, “Bitrate modeling of
scalable videos using quantization parameter, frame rate and spatial resolution,” in Proc. IEEE ICASSP, Dallas, TX, USA, Mar.
2010.
[35] S. Parakh and A. Jagannatham, “Game theory based dynamic
bit-rate adaptation for H.264 scalable video transmission in 4G
wireless systems,” in Proc. SPCOM, Bangalore, India, Jul. 2012.
[36] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge,
U.K.: Cambridge University Press, 2004.
[37] A. Takayama, Analytical Methods in Economics. Ann Arbor, MI,
USA: University of Michigan Press, 1993.
[38] A. Cambini and L. Martein, Generalized Convexity and Optimization:
Theory and Applications. Guildford, U.K.: Springer-Verlag, 2009.
[39] J. Reichel, H. Schwarz, and M. Wien, Joint Scalable Video Model
JSVM-12 Text, Joint Video Team (JVT) of ISO/IEC MPEG & ITU-T
VCEG Doc. JVT-Y202, Oct. 2007.
Chetna Singhal (S’13) received the B.Eng.
degree in electronics and telecommunications
from the Univ. of Pune in 2008 and M.Tech.
degree in computer technology from Electrical
Eng. Dept., IIT Delhi, in 2010. She was with
IBM Software Lab, New Delhi, as a Software
Engineer, from June 2010 to July 2011. She
is currently pursuing the Ph.D. degree from the
Bharti School of Telecommunications, IIT Delhi,
since July 2011. Her current research interests
include handoff schemes and cross-layer optimization in wireless networks, adaptive multimedia multicast and broadcast schemes and technologies. She is a student member of the ACM,
IEEE, and IEEE Computer and Communications Societies.
SINGHAL ET AL.: JOINT OPTIMIZATION OF USER-EXPERIENCE AND ENERGY-EFFICIENCY IN WIRELESS MULTIMEDIA BROADCAST
Swades De (S’02-M’04) received the Ph.D.
degree in electrical eng. from the State Univ.
of New York at Buffalo in 2004. He is currently
an Associate Professor of Electrical Eng. at IIT
Delhi. His current research interests include performance study, resource efficiency in wireless
networks, broadband wireless access, and communication and systems issues in optical networks. Dr. De currently serves as an Associate
Editor of IEEE Communications Letters and
Springer Photonic Network Communications
journal. He is a member of IEEE, IEEE ComSoc, and IEICE.
Ramona Trestian (S’08-M’12) received her
B.Eng. degree in telecommunications from
the Electronics, Telecommunications and the
Technology of Information Dept., Technical Univ.
of Cluj-Napoca, Romania, in 2007 and the Ph.D.
degree from Dublin City Univ., Ireland, in March
2012. She is a Lecturer with the Computer and
Communications Eng. Dept., School of Science
and Technology, Middlesex Univ., London, U.K.
She has published in prestigious international
conferences and journals and has two edited
books. She is a reviewer for international journals and conferences and
an IEEE member. Her current research interests include mobile and
wireless communications, multimedia streaming, handover and network
selection strategies, and software-defined networks.
1535
Gabriel-Miro Muntean (S’02-M’04) received the
B.Eng. and M.Sc. degrees in software eng. from
the Computer Sc. Dept., Politehnica Univ. of
Timisoara, Timisoara, Romania, in 1996 and
1997, respectively, and the Ph.D. degree from
the School of Electronic Eng., Dublin City Univ.
(DCU), Dublin, Ireland, in 2003 for his research
on quality-oriented adaptive multimedia streaming over wired networks. He is currently a Senior
Lecturer with the School of Electronic Eng., DCU.
He is Co-Director of the Performance Eng. Lab.
Research Group, DCU, and Director of the Network Innovations Centre,
part of the Rince Institute Ireland. He has published over 150 papers in
prestigious international journals and conferences, has authored three
books and 12 book chapters, and has edited five other books. His
current research interests include quality-oriented and performancerelated issues of adaptive multimedia delivery, performance of wired and
wireless communications, energy-aware networking, and personalized
e-learning. Dr. Muntean is an Associate Editor of the IEEE Transactions
on Broadcasting, the IEEE Communication Surveys and Tutorials, and
a reviewer for other important international journals, conferences, and
funding agencies. He is a member of the ACM, IEEE and the IEEE
Broadcast Technology Society.
For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.