Energy-efficient HTTP Adaptive Streaming for High- Quality Video over HetNets

Abstract

In this work, we propose an energy-efficient HTTP adaptive streaming for high-quality video over heterogeneous networks (HetNets). To support high-quality video and overcome the limitations of a single network, the proposed system downloads the video segments by concurrently using the HetNets. In the proposed system, the segment bitrate, number of requested segments, network sleep time, and size of requested data through each wireless network are determined adaptively to provide seamless high-quality video streaming in an energy and cost efficient way. The proposed system is fully implemented in an Android-based mobile device and tested in an actual wireless network environments.

Full text

Energy-efficient HTTP Adaptive Streaming for High-
Quality Video over HetNets
Yunmin Go1, Oh Chan Kwon2, and Hwanjun Song3
1Div. of IT Convergence Engineering, POSTECH (Pohang University of Science and Technology), Pohang, Republic of Korea
2Software R&D Center, Samsung Electronics, Suwon, Republic of Korea
3Dep. of Computer Science Engineering, POSTECH (Pohang University of Science and Technology), Pohang, Republic of Korea
gnfservant@postech.ac.kr, ochan.kwon@samsung.com, hwangjun@postech.ac.kr
Abstract—In this work, we propose an energy-efficient HTTP
adaptive streaming for high-quality video over heterogeneous
networks (HetNets). To support high-quality video and overcome
the limitations of a single network, the proposed system
downloads the video segments by concurrently using the HetNets.
In the proposed system, the segment bitrate, number of requested
segments, network sleep time, and size of requested data through
each wireless network are determined adaptively to provide
seamless high-quality video streaming in an energy and cost
efficient way. The proposed system is fully implemented in an
Android-based mobile device and tested in an actual wireless
network environments.
Keywords—HTTP adaptive streaming, Heterogeneous networks,
Multipath, Energy-efficient, Cost constraint
I. INTRODUCTION
Recently, HTTP adaptive streaming has been widely
deployed in popular video services such as YouTube and
Dailymotion. Moreover, MPEG Dynamic Adaptive Streaming
over HTTP (DASH) standard has been released in April 2012
[1]. In HTTP adaptive streaming, the server stores the video
data that are encoded at different bitrates and divided into
multiple fixed-length segments. The server records the segment
metadata, called media presentation description (MPD) in
MPEG-DASH. When a client requests the streaming service,
the client firstly receives the MPD from the server. Based on
the MPD, the client requests a proper segment according to the
network conditions. The network conditions, such as network
throughput and delay, are estimated while the client receives
the segments [2].
However, buffer underflow and low quality continuation
problem can still be occurred even in HTTP adaptive streaming.
Especially, the limitations of a single wireless network, e.g.
bandwidth limitation and time-varying wireless channel, and
user mobility, can severely degrade the performance of the
HTTP adaptive streaming. To solve this problem, many
researches have been performed to improve the network
throughput and quality of service (QoS) by simultaneously
harmonizing the HetNets [3]. Nonetheless, there are several
challenges to harmonize HetNets for HTTP adaptive streaming
services in mobile devices. First challenge is that much more
energy is consumed when the multiple network interfaces are
enabled. Second challenge is the monetary cost according to
the network usage, such as 3G and LTE. In general, a user may
prefer to receive a video streaming service within his/her own
cost plan. Third challenge is segment scheduling. When a
video segment is received through multiple wireless networks
with different characteristics, buffer underflow may occur due
to the out-of-order packet delivery. So far, many previous
works are proposed in order to defeat these challenges [3]-[9].
In this work, we propose an energy-efficient HTTP
adaptive streaming system for the provision of seamless high-
quality video streaming services over HetNets. The proposed
system dynamically adjusts the segment bitrate, number of
requested segments, network sleep time, and size of the
requested data through HetNets considering the energy
consumption and monetary cost constraints in mobile device.
According to our knowledge, there are no research studies on
HTTP adaptive streaming over HetNets while considering the
energy consumption, video quality, and user monetary cost
plan. The rest of this paper is organized as follows. In section II,
details of the proposed system are presented. Experimental
results are provided in section III, and concluding remarks are
given in section IV.
II. PROPOSED HTTP ADAPTIVE STREAMING SYSTEM
Based on MPEG-DASH standard, the proposed system
requests MPD from the web server through the wireless
network with strongest signal strength among the available
HetNets. Then the web server sends back the MPD to the client,
and the client analyzes the received MPD. With the analyzed
MPD, buffered video time ( buf
t), RTT ( rtt
t
), and network
throughput ( r

), the proposed system determines the optimal
control parameters such as the segment bitrate (
s
eg
v), number
of requested segments (
s
eg
n), segment duration time ( dur
t)
(including network sleep time), and segment scheduling map
(S) (i.e., the size of requested data through each wireless
network). Segments are requested from the web server
according to the determined control parameters. While the
client receives the segment from the server, the RTT and
throughput of each network is estimated. The received packets
are sequenced according to the segment in the reordering
buffer, and ordered data are delivered to the video decoder for
video playback.
In the proposed system, the segment duration plays as a
basic operation time unit as shown in Fig. 1. The segment
duration composed of three stages including the request stage
(including RTT of request message), segment download stage,
2015 IEEE 26th International Symposium on Personal, Indoor and Mobile Radio Communications - (PIMRC): Services
Applications and Business
978-1-4673-6782-0/15/$31.00 ©2015 IEEE 2262

and sleep stage. For a segment duration, multiple partitioned
segments can be requested by using byte-range request and
HTTP pipelining, and they have same segment quality in order
to provide consistent video quality to the client. In the proposed
system, energy consumption and cost are controlled by jointly
adjusting the requested data size through each wireless network
and sleep time of each wireless network. The sleep time is
determined considering the heterogeneous power states of the
wireless networks.
A. Problem Description
It is important to estimate the condition of each network in
order to efficiently utilize the HetNets. The proposed system
estimates the network condition by using the observed TCP
information such as TCP window size and RTT. The estimated
throughput of i-th network for the next segment duration i
r
 is
calculated by using an exponential weighted moving average
with the current TCP throughput and previous estimated
throughput. The segment scheduling map S is represented by
()
12
, ,..., seg
n
ss s=S  ,
()
1, 2, ,
, ,..., net
T
jjjNj
sss s=
,
where
j
s
 is the segment scheduling vector of the j-th
segment, ,ij
s
is the number of packets of the j-th segment
delivered through the i-th network, and net
N denotes the
number of usable wireless networks at a mobile device. In fact,
partitioned segment is requested as byte unit, but it causes high
computational complexity to obtain the optimal segment
scheduling vector. Thus, we assume that each packet includes a
fixed amount of bytes, and the number of packets is employed
instead of the number of bytes in order to reduce the
computational complexity. ,
s
eg
s
eg
j
v
g is the number of packets for
the j-th segment with segment bitrate
s
eg
v that is obtained by
,,
,1
net
seg
seg
N
seg
j
vpkt ij
jv i
g
bB s
=
==

, (1)
where ,
s
eg
j
v
bis the amount of bytes of the j-th segment when
the selected segment bitrate is
s
eg
v, and
p
kt
B is packet size. The
receive time of the requested segments through the i-th
network can be approximately calculated by
()
()
,
1
,seg
n
recv
iseg pkt iji
j
tn B sr
=
=⋅

S. (2)
The monetary cost vector c
 is symbolized by
()
12
, ,..., net
N
ccc c=
,
where i
c denotes the required monetary cost per packet for the
i-th network. The monetary cost for the segment duration
()
,
seg
nc n S is calculated by
()
1
,seg
n
s
eg j
j
nc n c s
=
=•

S
. (3)
Now, we can formulate our problem as follows.
Problem Formulation: Determine
s
eg
v,
s
eg
n, dur
t, and S for
the next segment duration to minimize the following cost
function
()
,,,
seg seg dur
vntΨS
()
()
()
,, 1 ,
s
eg dur seg seg
en t dn v
ωω
⋅+−⋅S
subject to
()
max
,
s
eg seg
nc n NC≤S, and (4)
()
{
}
()
1max , 1
net
rtt recv
ii bufseg
iN tt j t T j
δ
≤≤ +≤⋅+⋅−S for 1
s
eg
j
n≤≤ , (5)
where
()
,,
seg dur
en t S denotes the consumed energy per segment
during the segment duration,
()
,
s
eg seg
dn v is the average
distortion of the requested segments when the segment bitrate
is
s
eg
v, and
s
eg
T is the video playback time of a segment.
ω
is
a real number
()
01
ω
≤≤ that is a system parameter to pursue
an effective trade-off between energy consumption and video
quality (
ω
can be selected by the user preference), max
s
eg
NC
denotes the maximum monetary cost per segment duration, and
δ
is a guard gain for the client receive buffer to consider the
estimation uncertainty, i.e., the inaccurate estimation of RTT
and the segment download time over time-varying HetNets
()
01
δ
≤≤. Eq. 4 denotes the monetary cost constraint, and
max
s
eg
NC is calculated by
max max 60 acc
s
eg seg seg mnt res
NC n T NC NC=⋅⋅ + , (6)
where max
mnt
NC is the maximum monetary cost for a minute of
video playback, and acc
res
NC is the accumulated residual
monetary cost for the next segment duration (i.e., obtained by
subtracting the actual monetary cost from the maximum
monetary cost of the current segment duration). Eq. 5 means
that the requested segment should arrive at the client before a
buffer underflow occurs. In the proposed system, it is assumed
that the rate-distortion model [10] of each segment is
embedded in MPD in order to estimate the average distortion at
the client, that is, the corresponding average distortion of
s
eg
nsegments is modeled by
() ()
{
}
1
,seg j
n
s
eg seg j seg seg
j
dn v v n
ε
γ
=
=⋅
, (7)
where
()
jj
γγ
+
∈ℜ and
[
]
()
1, 0
jj
εε
∈− are model parameters
of the j-th segment depending on the video sequence.
B. Energy Consumption Model
We assume that the energy consumption patterns of all
wireless networks can be simplified as shown in Fig. 2. In this
work, we only considers the segment download stage (receive
state) and sleep stage (tail state and idle state) in a segment
duration. The energy consumption of the segment request
message is ignored since the size of segment request message
is very small (less than 1 KB) and RTT is too short (less than
100 ms). As shown in the figure, the sleep time
s
lp
i
t can be
calculated by
()
,
slp rtt recv
iduriiseg
ttttn=−− S. (8)
The tail time tail
i
t is determined according to
s
lp
i
t as follows.
if ,
otherwise,
tail slp tail
tail i i i
islp
i
TtT
tt
≥
=
 (9)
1Mbp s, 2,
30 30
4.5sec, 50 50
seg seg
dur
vn
t
==




==




S
Fig. 1. Example of segment duration when two wireless networks are
enabled (numbers in parenthesis refer to segment bitrate).
2015 IEEE 26th International Symposium on Personal, Indoor and Mobile Radio Communications - (PIMRC): Services
Applications and Business
2263

where tail
i
T is the fixed tail time of the i-th network [9]. idle
i
t is
the time in the idle state after tail
i
t, denoted by
- if ,
0 otherwise.
s
lp tail slp tail
idle ii i i
i
tt t T
t≥
=
 (10)
The power states of the i-th network are represented by
receive power recv
i
p
, tail power tail
i
p
, and idle power idle
i
p
,
respectively. recv
i
p
is obtained by linear model [9] of network
throughput, i.e.,
recv
iiii
pr
ημ
=⋅+
, (11)
where i
η
and i
μ
are model parameters that are obtained from
the energy profile on the corresponding wireless network
interface. Therefore, the consumed energy per segment at the
mobile device can be calculated by
() ()
{
1
1
,, ,
net
N
recv recv tail tail
seg dur i i seg i i
seg i
en t p t n p t
n=

=⋅+⋅


SS
}
()
{
}
1min
net
idle idle rtt
ii basedur i
iN
pt p t t
≤≤

+⋅ + ⋅−

, (12)
where base
p
is the base power of the mobile device.
C. Parameter Selection Algorithm
In this section, we propose the determining process of
optimal control parameters. In the proposed algorithm, some
candidates having no possibility of being selected as the
optimal solution are removed. Then optimal solution that
minimizes the given cost function is searched by performing a
full search among all the remaining candidates. The details of
the proposed algorithm are presented below.
Step 0) Initialize cur
s
eg
v, cur
s
eg
n
, cur
dur
t, cur
S, and
()
,,,
seg seg dur
vntΨ=∞S.
Step 1) Sort the wireless networks in ascending order
according to the receive power strength.
Step 2) Generate the combination of
()
,
new new
seg seg
vn for
new
s
eg seg
vV∈and max
1new
s
eg seg
nN≤≤ (where max
s
eg
N is the
maximum number of requested segments during the
segment duration, and
s
eg
V is the set of available
segment bitrates provided by the web server). For the
generated combinations, calculate the maximum
number of packets for the i-th network.
() ()
{
max min ,
new new
iseg buf seg seg seg i pkt
gn tnT TrB
δ

=+⋅−⋅ ⋅


}
max for 1 net
seg i iN
NC c ≤≤

 , (13)
If
()
max
,
11
new
net seg
new
Nn
new seg
iseg jq
ij
gn g
==
<

, then eliminate the
corresponding element
()
,
new new
seg seg
vn .
Step 3) Produce the combination of
()
,,
new new new
seg seg
vnS for the
filtered set
()
,
new new
seg seg
vn by Step 2. To determine new
S,
the possible range of i
s
that denotes the number of
packets through the i-th network is determined by
()
max
0 for 1
new
ii seg net
g
gn iN≤≤ ≤≤ . (14)
If i
g
is given, ,
new
ij
s
in new
S is attained by
()
,1
1
,,
1
1
,
,1
if and ,
if and ,
otherwise. (15)
net
new
seg
new
seg
Nnew
seg
net seg
ik
jr k
j
new new new
ij i ik net seg
k
i
seg new
kj
jr k
iN jngg g
sg s iN jn
gs
=
−
=
−
=

<<⋅



=− < =

−





Step 4) Select one from the combination set acquired by Step 3,
and calculate new
dur
t based on new
s
eg
n and new
S as follows.
()
{
}
()
max
1max , if ,
otherwise, (16)
net
rtt recv new new
i i seg buf buf
new iN
rnd new
buf seg seg
ttn t T
t
tn T
δ
≤≤
+≤

=
+−⋅


S
new
buf buf seg seg
ttnT=+ ⋅
, (17)
where buf
t
 is the estimated buffered video time after
the download of new
s
eg
n segments, and max
buf
T is the
maximum buffered video time. Now, calculate
()
,,,
new new new new
seg seg dur
vntΨS. If
()
,,,
new new new new
seg seg dur
vntΨS is
less than
()
,,,
cur cur cur cur
seg seg dur
vntΨS, then replace
()
,,,
cur cur cur cur
seg seg dur
vntS with
()
,,,
new new new new
seg seg dur
vntS.
Step 5) If all possible combinations of
()
,,
new new new
seg seg
vnS are
examined, then terminate the process with the optimal
solution
()
,,,
cur cur cur cur
seg seg dur
vntS. If not, then go back to
Step 4.
In general,
s
eg
V and net
N are typically a small integer
because the streaming service provider cannot support fine-
grained bitrates and state-of-the-art smart mobile devices only
has a Wi-Fi interface and a cellular network interface. In this
case, the required computational complexity of the proposed
algorithm is
()
()
()
max max max
12
max
seg seg i seg
i
VN g N
O≤≤
⋅⋅ . It is low to
operate in real-time at mobile device.
III. EXPERIMENTAL RESULTS
During the experiment, the proposed system is
implemented to support MPEG-DASH [1] at Samsung Galaxy
S4 LTE-A. In addition, MPD and segments also are created on
the basis of the MPEG-DASH. We use full HD video that are
Big Buck Bunny, Red Bull Playstreets, and The Swiss Account
[11]. All videos are encoded by H.264/MPEG-4 AVC at five
different bitrates (Big Buck Bunny: 1, 2, 3, 5, and 8 Mbps, and
the others: 1, 2, 3, 5, and 6 Mbps). Total video playback time is
120 seconds, and the initial video data buffering time is fixed
to 4 seconds (i.e., two segments). Details of the experimental
parameters are summarized in Table I.
The proposed system is examined through Wi-Fi and LTE
network simultaneously. To emulate the time-varying wireless
network conditions, DummyNet [12] is employed at the Wi-Fi
Fig. 2. Simplified energy consumption model.
2015 IEEE 26th International Symposium on Personal, Indoor and Mobile Radio Communications - (PIMRC): Services
Applications and Business
2264

access point. During the experiment, the RTT of Wi-Fi path is
fixed to 40 ms, and network throughput of Wi-Fi path is set to
change every 20 seconds (6 Mbps → 2 Mbps → 4 Mbps → 2
Mbps → 6 Mbps). On the other hand, the condition of LTE
network cannot be adjusted since it is managed by a
commercial telecommunication company. The measured end-
to-end throughput of LTE network is approximately 20 Mbps,
and the measured RTT is about 58 ms. To measure the
consumed energy, the Monsoon Power Monitor is used. The
parameters of the energy consumption model are empirically
measured as shown in Table II, and embedded in the MPD.
The required monetary cost per packet for Wi-Fi and LTE are
set to 0 Korean Won (KRW) and 0.0285 KRW, respectively.
A. Performance Verification of Energy Consumption Model
We verify the energy consumption model with the
measured energy consumption data. Fig. 3 presents the
comparison between measured and estimated energy
consumption for each segment duration. It is apparently
observed that the estimated energy fits well with the measured
energy. However, the estimation error between the measured
energy and estimated energy still exists due to unpredictable
job processing in the operating system, and inaccurate
estimation of network condition and processing power
consumption. The average estimation error rate of both Wi-Fi
and LTE are approximately 18.25%.
B. Performance according to Control Parameters
We examine the performance of the proposed system with
regard to the segment bitrate and the number of requested
segments. For the experiment
ω
is fixed to 0.05, and test video
is Big Buck Bunny. As indicated in Table III,
s
eg
vaffects
energy consumption because the segment data size increases as
s
eg
v increases, and more energy is required to receive higher
bitrate segments. The throughput increases as
s
eg
n increases
because the larger
s
eg
n provides enough time to increase the
TCP window size. However, the energy consumption also
increases because the download time for the requested segment
increases. When
s
eg
v is fixed to 8 Mbps, the bitrate of the
requested segment is so large that the network throughput may
be significantly changed during the segment download, which
sometimes causes client buffer underflows. Consequently, the
proposed system adjusts
s
eg
v and
s
eg
n dynamically to minimize
the cost function and avoid buffer underflows.
C. Performance according to
ω
In this section, we investigate the performance of the
proposed system according to
ω
. The maximum monetary
cost per selected video (120 seconds) is set to 350 KRW, and
test video is Big Buck Bunny. Fig. 4 presents the performance
of the proposed system when
ω
is 0.90. The proposed system
provides relatively low video quality as shown in Fig. 4 (a),
and stays in the sleep state for a longer time to save the energy
as shown in Fig. 4 (b). In this case, the energy consumption
plays a more important role than the monetary cost, and the
proposed system consumes about 9,946 μAh and 321.6 KRW,
which is less than the maximum monetary cost. As illustrated
in Fig. 5, when
ω
is set to 0.10, the proposed system fully
utilizes the LTE network with the maximum monetary cost
constraint, i.e., the expended monetary cost is 345.7 KRW and
the consumed energy is 14,560 μAh. The summarized results
are presented in Table V.
Table I. Parameter values used in the experiment.
Parameter Description Value
net
N Number of usable wireless networks at a mobile device 2
s
e
g
V Number of segment quality levels for the requested video 5
max
s
e
g
N Maximum number of requested segments 4
max
mnt
N Maximum cost per minute of video playback (KRW) 175
β
Weighting factor for network throughput estimation 0.25
δ
Client receive buffer parameter 0.90
p
kt
B Packet size (byte) 1460
max
bu
f
T Maximum buffered video time (seconds) 20
s
e
g
T Video display time of a segment (seconds) 2
Table II. Power profile of the Wi-Fi and LTE.
Parameter Wi-Fi LTE
Receive power
η
=210.67(mW/Mbps),
μ
=374.91(mW)
η
=132.32(mW/Mbps),
μ
=1433.90(mW)
Tail power 310.66 (mW) 657.41 (mW)
Idle power 24.93 (mW) 26.20 (mW)
Tail time (fixed) 0.2 (seconds) 10 (seconds)
※ Base power = 774.27 (mW)
020 40 60 80 100 120
0
500
1000
1500
2000
Time (sec.)
Consumed energy (uAh)
Measured
Estimated
Fig. 3. Energy consumption pattern when both Wi-Fi and LTE are utilized.
020 40 60 80 100 120
0
2000
4000
6000
8000
10000
Time (sec.)
Segment bitrate (Kbps)
020 40 60 80 100 120
LTE
Wi-Fi
Time (sec.)
(a) (b)
020 40 60 80 100 120
0
1
2
3
4
5
Time (sec.)
n
seg
010 20 30 40 50 60
0
50
100
150
200
250
300
350
Segment number
Networking cost (KRW)
Consumed networking cost
Maximum networking cost
(c) (d)
Fig. 4. Performance of the proposed system when 0.10
ω
=: (a) requested
segment bitrate, (b) network state (grey region is sleep state), (c) number of
requested segments, and (d) cumulative consumed monetary cost.
Table III. Performance comparison with fixed vseg and nseg.
v
seg
(Mbps) nseg Energy
(μAh)
Average
PSNR(dB)
Throughp
ut (Kbps)
Buffering
Time (sec.)
3
1 9523 40.58 4285 0
2 10234 40.58 4769 0
4 11654 40.58 5242 0
5
1 15036 42.38 4629 15.27
2 15912 42.38 4690 18.13
4 16222 42.38 5145 23.23
Proposed Alg. 13456 42.12 4745 0
2015 IEEE 26th International Symposium on Personal, Indoor and Mobile Radio Communications - (PIMRC): Services
Applications and Business
2265

D. Performance Comparison with Existing Algorithms
In this section, the performance of the proposed algorithm
is compared with existing system that are Dynamic sub-
segment approach (SubSeg) [7] and GreenBag [8] that are
modified slightly for our experiment. For a fair comparison,
GreenBag employs the quality adaptation algorithm of SubSeg
because it has no quality adaptation algorithm. In addition,
SubSeg and GreenBag employ the same monetary cost strategy
of the proposed algorithm (see step 2 in Section III.C).
ω
of
the proposed algorithm is set to 0.20. The test videos are Big
Buck Bunny, Red Bull Playstreets, and The Swiss Account.
The experiment results are summarized in Table V. It is clearly
observed that the proposed system consumes the lowest energy
since it determines not only the amounts of requested segment
data but also the sleep time of the networks in order to save
energy. In addition, the proposed system exhibits the highest
video quality because it considers the uncertainty of RTT and
segment download time over time-varying HetNets. Therefore,
the proposed algorithm shows best performance in terms of
energy consumption and video quality for all the video
sequences.
IV. CONCLUSION
In this work, we have proposed an energy-efficient HTTP
adaptive streaming system that provides seamless high-quality
video streaming services with cost constraints over HetNets. In
the proposed system, the control parameters such as segment
bitrate, number of requested segments, segment duration, and
segment scheduling map are dynamically determined to pursue
an effective trade-off among video quality and energy
consumption. The experimental results shows that the proposed
system can provide high-quality HTTP adaptive streaming
services over HetNets while reducing the energy consumption
of mobile device and satisfying the user monetary cost plan.
ACKNOWLEDGMENT
This work was partly supported by the ICT R&D program
of MSIP/IITP [13-911-04-005, Research and Development of
5G Mobile Communications Technologies using CCN-based
Multi-dimensional Scalability] and Basic Science Research
Program through the National Research Foundation of
Korea(NRF) funded by the Ministry of Education (NRF-
2013R1A1A2006732).
REFERENCES
[1] ISO/IEC 23009-1, “Information technology – Dynamic adaptive
streaming over HTTP (DASH) – Part 1: Media presentation description
and segment formats,” Apr. 2012.
[2] T. Stockhammer, “Dynamic adaptive streaming over HTTP: standards
and design principles,” in Proceedings of the ACM conference on
Multimedia Systems (MMSys), pp. 133-144, Feb. 2011.
[3] A. L. Ramaboli, O. E. Falowo, and A. H. Chan, “Bandwidth aggregation
in heterogeneous wireless networks: A survey of current approaches and
issues,” Journal of Network and Computer Applications, vol. 35, no. 6,
May. 2012.
[4] M. A. Hoque, M. Siekkinen, and J. K. Nurminen, “On the energy
efficiency of proxy-based traffic shaping for mobile audio streaming,” in
Proceedings of the IEEE Consumer Communications and Networking
Conference (CCNC), pp. 891-895, Jan. 2011.
[5] M. A. Hoque, M. Siekkinen, and J. K. Nurminen, “TCP receive buffer
aware wireless multimedia streaming: an energy efficient approach,” in
Proceeding of the ACM Workshop on Network and Operating Systems
Support for Digital Audio and Video (NOSSDAV), pp. 13-18, Feb. 2013.
[6] R. Ding and G.-M. Muntean, “Device characteristics-based
differentiated Energy-efficient Adaptive Solution for video delivery over
heterogeneous wireless networks,” in Proceedings of the IEEE Wireless
Communications and Networking Conference (WCNC), pp. 4588-4593,
Apr. 2013.
[7] D. H. Bui, K. Lee, S. Oh, I. Shin, H. Shin, and H. Woo, “GreenBag:
Energy-efficient Bandwidth Aggregation for Real-time Streaming in
Heterogeneous Mobile Wireless Networks,” in Proceedings of the IEEE
Real-Time Systems Symposium (RTSS), pp. 57-67, Dec. 2013.
[8] K. Evensen, D. Kaspar, C. Griwodz, P. Halvorsen, A. F. Hansen, and P.
Engelstad, “Improving the performance of quality-adaptive video
streaming over heterogeneous access networks,” in Proc. ACM Conf.
Multimedia Systems (MMSys), 2011, pp. 57-68.
[9] J. Huang, F. Qian, A. Gerber, Z. M. Mao, S. Sen, and O. Spatscheck, “A
close examination of performance and power characteristics of 4G LTE
networks,” in Proceedings of the International conference on Mobile
Systems, applications, and services (MobiSys), pp. 225-238, Jun. 2012.
[10] D. Jurca and P. Frossard, “Media flow rate allocation in multipath
networks,” IEEE Transactions on Multimedia, vol. 9, no. 6, pp. 1227-
1240, 2007.
[11] ITEC – Dynamic Adaptive Streaming over HTTP, http://www.itec.uni-
klu.ac.at/dash.
[12] L. Rizzo, “Dummynet: a simple approach to the evaluation of network
protocols,” ACM SIGCOMM Computer Communication Review, vol.
27, no. 1, pp. 31-41, Jan. 1997.
Table IV. Performance comparison with existing algorithms.
Video Algorithm Energy (μAh) Average PSNR (dB)
Big Buck
Bunny
SubSeg 13319.24 41.56
GreenBag 12544.31 41.89
Proposed 11160.51 42.04
Red Bull
Playstreets
SubSeg 12599.70 40.15
GreenBag 12525.62 40.48
Proposed 10927.62 41.56
The Swiss
Account
SubSeg 10208.85 40.57
GreenBag 9854.95 40.78
Proposed 9657.60 41.60
020 40 60 80 100 120
0
2000
4000
6000
8000
10000
Time (sec.)
Segment bitrate (Kbps)
020 40 60 80 100 120
LTE
Wi-Fi
Time (sec.)
(a) (b)
020 40 60 80 100 120
0
1
2
3
4
5
Time (sec.)
n
seg
010 20 30 40 50 60
0
50
100
150
200
250
300
350
Segment number
Networking cost (KRW)
Consumed networking cost
Maximum networking cost
(c) (d)
Fig. 5. Performance of the proposed system when 0.90
ω
=: (a) requested
segment bitrate, (b) network state (grey region is sleep state), (c) number of
requested segments, and (d) cumulative consumed monetary cost.
Table V. Performance of the proposed system according to
ω
.
ω
Energy
(μAh)
Average PSNR
(dB)
Networking cost
(KRW)
0.90 9946.27 40.48 321.64
0.50 11037.00 40.92 342.89
0.10 14560.00 42.10 345.70
2015 IEEE 26th International Symposium on Personal, Indoor and Mobile Radio Communications - (PIMRC): Services
Applications and Business
2266