QOE CONTINUUM DRIVEN HTTP ADAPTIVE STREAMING OVER MULTI-CLIENT

WIRELESS NETWORKS

Zhisheng Yan, Jingteng Xue, and Chang Wen Chen

The State University of New York at Buffalo

Buffalo, NY 14260, USA
{zyan3, jingteng, chencw}@buffalo.edu

ABSTRACT the video with the highest possible quality while the band-
width can also be efficiently utilized.
Different from traditional HTTP adaptive streaming (HAS)
One key objective of HAS is to improve the Quality of
in which only one client is considered, HAS over multi-client
Experience (QoE) of users under highly complicated wire-
wireless networks faces new challenges. The Quality of Ex-
less environment. One major challenge is how to precisely
perience (QoE) of users becomes unstable due to users’ com-
measure users’ QoE, which is a subjective perception of the
petition for shared bandwidth. It is thus important to accu-
entire viewing experience. Users’ QoE at a certain moment is
rately estimate the perceived experience of users and then
impacted by the playback quality of both currently displayed
adapt the streaming process accordingly. Furthermore, the
frame and previously displayed frames, as well as the con-
QoE fairness among multiple clients subscribing to the same
sistency in playback quality [1]. Conventional modeling of
services shall also be addressed. In this research, we pro-
QoE simply based on current viewing experience makes HAS
pose a QoE continuum driven HAS adaptation algorithm to
adaptation non-optimal. Therefore, QoE measurement shall
address these challenges. We model the QoE continuum as an
consider users’ viewing experience in a temporally continu-
integrated consideration of cumulative playback quality and
ous manner. Furthermore, the QoE fairness among different
playback smoothness. Based on this model, we jointly opti-
HAS users who have the same service priority should also be
mize the quality adaptation of multiple users by considering
guaranteed. According to [2], the time-varying and shared
both QoE history and channel status. Moreover, we propose
wireless networks lead to unpredictable QoE for every user.
to use quantization parameter and segment size to represent
Consequently, users may have very different viewing experi-
the video files in a fine-grained fashion, in order to more ef-
ence due to their different channel status even though they
fectively capture the bandwidth fluctuation. The results from
pay for the same service. Another important issue in cur-
extensive simulations show that the proposed scheme can pro-
rent HAS systems is the inaccurate representation of the pre-
vide balanced and satisfactory QoE among multiple clients.
encoded video files. Currently, different video quality lev-
Index Terms— QoE continuum, QoE fairness, HTTP els are usually indicated by video bit-rate. However, since
adaptive streaming, multi-client, wireless networks modern videos are all encoded in a variable-bit-rate (VBR)
fashion, the real bit-rate of different video segments is sig-
1. INTRODUCTION nificantly different and cannot be accurately represented by
an average bit-rate. To overcome all these adversaries, it is
With the development of powerful smart phones and tablets,
imperative to develop a new quality adaptation scheme that
and the growing demand of watching videos from anywhere
incorporates proper QoE modeling and adaptation measures.
and at anytime, video streaming over wireless networks has
been rapidly booming in the past few years. It is predicted that 1.1. Existing General HAS
video traffic will account for over two-thirds of total mobile Recently, both 3GPP and MPEG have made tremendous ef-
traffic by the end of 2017. Lately, HTTP adaptive streaming forts towards the standardization of HAS [3]. However, spe-
(HAS) has been widely studied to address the bandwidth in- cific adaptation strategies are not part of the standard and are
efficiency in traditional streaming systems. The video source left to future designs. The overview of standardized HAS
is pre-encoded in several quality levels and is split into small QoE metrics and QoE-driven adaptation is presented in [3].
segments. The client dynamically requests the video segment Although several commercial HAS solutions, such as Mi-
with different quality at each switching point based on its net- crosoft Smooth Streaming and Apple Live Streaming, have
work and device status. That way, the user is able to watch been deployed, experimental results showed that the user ex-
This research is supported by NSF Grant 0915842 and Gift Funding from perience is negatively impacted when multiple clients com-
Huawei Technologies, Inc. pete for the shared wireless bandwidth [4]. Research com-
 QP11,S11 ... QPX1,SX1
munities have also proposed several HAS rate adaptation al- Xsegments
... QPX2,SX2
withMlevels QP12,S12
gorithms [5–7]. However, these algorithms are targeted at ... ... ...
QP1M,S1M ... QPXM,SXM
single-user and client-side adaptation. They cannot be di- HASserver
Router Router

rectly applied to the multi-client wireless networks since they UE1

are unaware of QoE fairness. The proposed framework, in- NUEs

UE2
GGSN
... SGSN
stead, shifts the adaptation to the base station that can jointly UEN
NodeB RNC
adapt the video quality and optimize the QoE of multiple
clients without modifying the standard HAS framework. Fig. 1. System Architecture
Only few work has been focused on HAS for multi-client base station only need to modify the clients’ requests to ac-
wireless networks. In [4], the authors first identified the is- commodate the QoE and fairness driven adaptation.
sues in multi-client wireless networks and proposed a simple The rest of this paper is organized as follows. In Section
traffic shaping mechanism to improve the experience of two 2, we describe the system models. In Section 3, the proposed
competing users. In [8], the authors enhanced the QoE by quality adaptation algorithm is introduced. We then evaluate
maximizing the overall mean opinion score (MOS) that is de- the performance of proposed algorithm and present the results
cided by the selected bit-rate and content of the video. Never- in Section 4. Finally we conclude this paper in Section 5.
theless, the QoE model is not accurate enough since only the
viewing experience at the adaptation moment is considered. 2. SYSTEM MODEL
2.1. System Architecture
1.2. Existing QoE-driven HAS
In this research, we consider the system architecture as shown
QoE has been studied in the design of several HAS systems.
in Fig. 1. The HAS server stores several VBR videos that all
QDASH [1] improved the HAS adaptation by incorporating
have M levels of quality. Each level of video is split into
a intermediate level into the switching process. However, no
multiple video segments with the same segment length and
explicit model is provided for QoE measurement, which lim-
each segment is characterized by the QP and segment file size,
its its application to different HAS systems. In [9], an adap-
i.e., a tuple (QP, S).
tation proxy located at the edge of the wired network was
The proposed algorithm is designed for cellular networks,
proposed to maximize user average data rate and minimize
where the base station jointly optimizes the quality adapta-
the rate variation and delay jitter. However, the algorithm
tion to ensure QoE continuum and fairness. Here we consider
is essentially QoS-driven and the standard HAS framework
3G High-Speed Downlink Packet Access (HSDPA) network
is modified due to the use of split-TCP. Although the work
as the underlying cellular network. However, the design prin-
in [10] revealed the importance of temporal factors for QoE
ciples are generic and the algorithm can be easily extended to
by studying the jitter and local content, the necessity of hu-
other cellular networks such as LTE. We focus on the HAS
man intervention makes it difficult to be generally deployed.
within one cell, where N users in the cell are managed by
1.3. Summary and Contributions Node-B and each user is indexed by i, i = 1, 2 · · · N . We
In summary, although aforementioned works have made con- assume that one user can only establish one flow with the
tributions to the HAS development, none of them thoroughly HAS server. The proposed algorithm is operated by Node-B
studies the impacts of QoE continuum (namely previous play- and implemented on top of underlying scheduling algorithms.
back quality, current playback quality, and playback quality Hence, it is practical to apply it to the modern cellular net-
variation) on HAS quality adaptation, and the QoE fairness works without modifying lower-layer scheduling strategies.
among multiple clients in wireless environment. The ma- The quality adaptation is proceeded as follows. Initially,
jor contribution of this research is that we propose a quality the HAS server sends out the media presentation description
adaptation algorithm that can guarantee both QoE and fair- so that Node-B and clients will have the knowledge of avail-
ness in one shared cell with multiple clients by exploiting able video representations. At each adaptation period whose
the nature of human perception and video source. Specif- length equals to segment length, clients request a specific
ically, we model the QoE continuum by considering both video segment at a certain quality level based on a simple
cumulative playback quality and playback smoothness. By throughput calculation, which requires very low complexity.
exploiting the proposed model, the base station can jointly Such operation is only used to be compatible with current
optimize the video quality levels of multiple HAS users un- DASH standards. Hence, no local hardware or operating sys-
der bandwidth-limited cellular networks, in order to fairly tem optimization is needed. Rather than directly forwarding
maximize all users’ QoE. Moreover, we propose to adopt the requests, however, Node-B in the proposed system will in-
fine-grained video representations that characterize the video tercept the requests and modify the adaptation decisions based
quality levels by a tuple of file size and quantization parame- on the proposed algorithm, where both low-layer link status
ter (QP), in order to capture video source characteristics and (such as channel quality indicator (CQI)) and high-layer QoE
then execute more efficient quality adaptation. More impor- and encoding information (such as current cumulative play-
tantly, the proposed algorithm is standard-friendly since the back quality) are utilized to guarantee balanced QoE contin-
uum for all users. Such information are embedded in the pe- where Ek is signified by the playback quality of the last dis-
riodic feedback from clients. Note that this is feasible since played frame j, i.e., Ek = qj,play . Besides, Lk represents the
3GPP DASH standard has standardized the quality metrics residual norm between expected frame and frame j in pixel
r
reporting process for clients. It uses a HTTP POST as the re- domain, which is linearly approximated by j,play η , where
porting protocol. Thus users are able to enjoy the video with rj,play is the bit count of the last displayed frame j and η
optimized QoE while neither the client nor the HAS server is is the compression ratio. To further describe the logarithmic
aware of the adaptation done by Node-B. Therefore, the pro- relation in rate distortion theory, we apply a logarithmic op-
posed algorithm can be friendly implemented in current HAS eration and then bound Lk in (0,1] as follows,
standard framework.
log(min(rj,play , rQP )) + c
Lk|j = − (4)
2.2. QoE Continuum Model log(rQP ) + c
We have identified that the QoE of HAS shall be measured in
a timely continuous fashion. In this section, we introduce two where rQP is the upper bound of frame size with the specific
factors that impacts QoE continuum, i.e., cumulative play- QP, and c = log(η)/2 is a model constant. The value of rQP
back quality (CPQ) and playback smoothness. can be calculated online based on previous streaming infor-
mation of decoded bits assuming that the size of frames with
2.2.1. Cumulative Playback Quality
a fixed QP is governed by a Laplace distribution. This as-
It has been discovered by psychological research that human sumption is reasonable given the independent identically dis-
memory demonstrates an exponential decay with respect to tributed property of compressed frames.
time, which is called as forgetting curve effect. Such effect
has been suggested by ITU standard [11] to be applied in 2.2.2. Playback Smoothness
continuous quality evaluation. Hence, we exploit this effect Subjective tests have shown that users prefer consistently
to model the CPQ. In [12], we have derived Qk , the CPQ at low-quality video over the video that fluctuates between high
frame k, as the summation of instantaneous playback quality quality and low quality [1]. By using the cumulative play-
over all displaying moments until the measure moment, i.e., back quality model in (1), the playback smoothness can be
effectively characterized in this research. Suppose one has
Qk = γQk−1 + (1 − γ)qk (1)
been enjoying the video with decent quality for a while (e.g.,
where Qk−1 is the CPQ at the previous frame, qk is the in- Qk−1 = 0.9), the sudden quality degradation of the current
stantaneous playback quality at frame k, and γ is the char- frame (e.g., qk = 0.6) could lead to a decreased current CPQ
acterization constant of the memory strength. Qk , qk and γ Qk . Such degradation would be accumulated and eventually
all belong to (0,1]. That way, we can capture the QoE from cause annoy experience after the playback of an entire seg-
previous displaying moments until the current moment. ment. The larger difference between the previous CPQ and
At a particular displaying moment of HAS systems, the the instantaneous playback quality, the worse the current CPQ
video player can either playback one frame normally or freeze would be. That way, the proposed algorithm places a con-
at a certain previous frame. Note that transmission distortion straint on the quality adaptation so that abrupt quality change
is disregarded in this research since HAS is virtually loss free is avoided and QoE continuum is improved. More impor-
due to the underlying TCP mechanisms. For the moment with tantly, due to the smooth quality variation of some users, re-
normal playback, the instantaneous playback quality is dic- sources can be saved for the other users who need them.
tated by the image quality of that frame. Thus we can predict
image quality from QP using a linear model and estimate the 3. PROPOSED QUALITY ADAPTATION
instantaneous playback quality qk,play as In this section, we introduce the novel framework to exploit
the nature of QoE continuum and fine-grained video repre-
qk,play = aQPk + b (2) sentations in order to fairly enhance the perceived experience
where a and b are content-specific parameters. of users. We formulate the adaptation optimization problem
When the bit-rate of the streamed video for a user ex- and propose an effective solution to achieve improved QoE
ceeds the user’s available bandwidth and the selected video continuum and QoE fairness.
has not been downshifted to a lower quality, playback inter- 3.1. Formulation of the Optimization
ruption may occur due to the client’s re-buffering. In this The QoE of HAS systems is critically decided by whether or
case, the video player’s screen will stall at the most recent dis- not the data volume of the streamed segment is larger than the
played image and consequently the user will undergo certain currently available bandwidth. It is necessary to incorporate
loss of expected visual information. Thereby, we can reliably the channel condition into quality adaptation process, espe-
model and validate the instantaneous playback quality for an cially considering the time-varying nature of wireless chan-
interruption moment qk,stall as the visual information loss L nel. In typical HSDPA implementations, wireless resources
scaled by the instantaneous user expectation E [12], i.e., are divided into Transmission Time Intervals during which
qk,stall = Lk Ek . (3) one user can receive its data packets. The maximum number
of data bits that can be received by user i per second, denoted The wisdom of adaptation behind (6) is that higher quality
as Ri,max , is essentially determined by the CQI of link i. At a level is generally given to those users who currently possess
given switching point, we employ the mean CQI of link i dur- a lower QoE continuum value and a better channel condition,
ing the last adaptation period to estimate Ri,max during the while significant quality variation shall also be avoided. For
next period by using the look-up table in 3GPP standard [13]. example, when users are currently enjoying the same level
Note that the mean CQI shall not be calculated based on a of video and the same channel condition, higher quality is
time interval that is too brief since it may not reflect the av- assigned to the user with a lower current QoE continuum
erage channel status in the next updating period. Similarly, because such adaptation will attain a maximum increase of
conservative mean CQI calculation using a long time interval the average QoE continuum. In other words, when a user
may lead to slow adaptation to channel variation. Thus the enjoys good experience for a long time, his/her satisfaction
resource sharing of user i, denoted by ϕi , is given by will rise less than the one with bad previous experience if the
video quality is raised. Consequently, we can enhance not
Si
ϕi = (5) only the QoE but also the fairness of users. Additionally, the
T Ri,max
penalty on playback variations avoids the sudden big change
where we approximate the bit-rate of the selected video seg- and keeps the playback smooth. This shall also inherently im-
ment as the ratio between the file size and the segment length. prove fairness since one’s potential resources for big quality
Hence, the available channel
 resource constraint for the pro- upgrading can be conserved for the others who need them.
posed HAS system is i∈N ϕi ≤ 1, where N is set of
HAS users. Such joint consideration of shared bandwidth will 3.2. Greedy Optimization
make the quality selection fair and reliable, and thus enhance The objective of (6) is nonlinear due to the involved loga-
QoE continuum. This definitely cannot be accomplished by rithmic and minimum operation, as well as the recursive cal-
individually blind client-side adaptation. culation process of Qi,t+T . Therefore, finding the optimal
According to the estimated maximum data rate, Node-B solution is complicated and time-consuming. We propose a
is able to estimate the CPQ at the next switching point Qi,t+T greedy algorithm, shown in Algorithm 1, to efficiently solve
when a video quality level li,t+T is considered. Node-B first the optimization and approximate the optimal solution. When
analyzes the adaptation-related information, such as current the algorithm initiates, Node-B collects users’ CPQ at t and
CPQ Qi,t and buffer status, as feedback received from clients.  at users’ current levels lt . At each
starts the greedy search
Then Qi,t+T can be recursively calculated according to (1) by subsequent step, if i∈N ϕi < 1, a small amount of re-
predicting whether the player is playing or stalling at each dis- sources that gain one-level quality improvement are assigned
play moment from t to t + T . If user i is not re-buffering at to the user who can accomplish maximum Δ i,in , the increase
t, Qi,t+T will be estimated from Qi,t by first considering the of average QoE continuum per unit data. If i∈N ϕi > 1, the
normal playback of the remaining frames (quality level li,t ) user having the lowest decrease of average QoE continuum
in the buffer and then considering the normal playback of the per unit data (Δi,de ) will be degraded one level. This pro-
selected level li,t+T for the rest of the period. If the user is cess repeats until all the resources are allocated or no further
re-buffering, Qi,t+T will be first calculated by assuming that change can be seen in average QoE continuum. The formu-
the player is frozen until the buffer size reaches the playback lated problem is a generalization of bounded knapsack prob-
threshold. The interruption time is decided by the estimated lem, which is NP-hard. The greedy heuristic is adopted to
Ri,max . Then Qi,t+T will be further updated by assuming solve the problem in polynomial time with O(M N logN ).
the normal playback of video with level li,t+T . That way,
we can accurately estimate the cumulative user experience at Algorithm 1 Greedy Quality Adaptation Algorithm
next switching point and assign the most satisfactory and fair 1: procedure A DAPT (Q, B, lt )  B:buffer status
adaptation decision to users accordingly. Note that interrup- 2: li,opt
 ← l i,t , ∀i ∈ N
tion may happen in reality during the playback of level li,t+T 3: if i ϕi <1 then  quality upgrading
due to the channel estimation error. 4: while i ϕi < 1 & objective in (6) changed do
Based on the above analysis, we now can formulate an op- 5: for i ∈ N do
timization to find the optimal level of video segment, which is 6: Update Δi,in , Δmax , maxue
indicated by a tuple (QP, S), for each user at switching point 7: lmaxue ,opt = min(lmaxue ,opt + 1, M )
t. We model the QoE continuum as a joint consideration of 8: else  quality degradation
the playback quality and playback smoothness, i.e., the CPQ 
9: while i ϕi ≥ 1 & objective in (6) changed do
model in (1). The objective of the optimization is to maximize 10: for i ∈ N do
the average QoE continuum of all users at the next switching 11: Update Δi,de , Δmin , minue
point t + T , subject to the wireless resources constraint. i.e.,
 12: lminue ,opt = max(lminue ,opt − 1, 1)
1
max(QP,S)  N i∈N Qi,t+T (6) 13: return lopt  i.e., li,t+T , i ∈ N
s. t. i∈N ϕi ≤ 1
 Table 1. Simulation Parameters Table 2. Playback Smoothness
l1 l2 l3 l4 l5 UE1/UE2
Metrics
Seg 1 bytes 13644 26952 52778 104951 141714 Baseline InstRate Proposed
Seg 2 bytes 13534 27166 53250 108734 148416 NoC 41/43 44/36 30/32
Seg 3 bytes 11066 53250 46748 43365 133706 PS 251.82/149.87 171.96/247.23 289.06/273.98
Seg 4 bytes 16027 31211 59114 3028 151288
Seg 5 bytes 25156 50353 98568 3284 253683
Table 3. Playback Quality
qplay 0.85 0.88 0.92 0.94 0.95
UE1/UE2
Metrics
4. PERFORMANCE EVALUATIONS Baseline InstRate Proposed
In this section, we compare the performance of the proposed APQ 4.09/3.84 4.03/4.39 4.35/4.26
algorithm with reference algorithms through simulations. We CPQ 0.94/0.81 0.89/0.93 0.93/0.93
first implement the algorithm in [4] (referred as Baseline) be- PoI 0/2.23 0/4.30 0/0
cause it is the first HAS adaptation scheme for multi-client NoI 0/1 0/2 0/0
wireless networks. We also implement a typical algorithm
algorithms, video playback of one user is usually unfairly
(referred as InstRate) that covers the logic behind many ex-
smoother than the other user.
isting works, in which the adaptation maximizes the utility
In order to demonstrate the quality adaptation wisdom be-
function dictated by selected instantaneous bit-rate, subject to
hind the proposed algorithm, we show the trend of channel
the channel constraint. One example of InstRate-like algo-
variation and quality level variation in Fig. 2. It is clear that
rithms is [8], where utility is the MOS mapped from bit-rate.
the proposed algorithm outperforms the reference algorithms
We focus on the architecture shown in Fig. 1. The HAS
with smoother quality change. Besides, the proposed scheme
server provides the test sequence “Stefan” with 5 levels of
does not simply capture the channel variation as is the case in
video whose QP is 47, 42, 37, 32, and 30 respectively. The
InstRate algorithm. Instead, the proposed algorithm is regu-
segment length T is 2 seconds and the frame rate is 30 fps.
lated by the smoothness penalty constraint and is less aggres-
The sequence has 300 frames and these 5 segments are re-
sive than reference algorithms, in which the highest possible
peatedly streamed. The file size of each segment for each
quality video is always assigned to users.
level l is shown in Table 1. We use EURANE for ns-2 [14] to
implement the underlying HSDPA network. We consider two 4.2. Playback Quality
users subscribing the same services but having different chan- We use CPQ after simulation ends and the average playback
nel status as in [4]. The typical wired and wireless network quality (APQ) inherited from [7] to evaluate playback qual-
parameters shown in [14] are used. Regarding the parame- ity viewed by users. APQ is defined as the weighted
P P sum of
ters of cumulative playback quality model, we inherit them the level index, i.e., APQ = p=1 (np × l)/ p=1 np . We
from [12], wherein the accuracy of the model is validated by also evaluate the playback quality by exploring the interrup-
both objective and subjective tests. The memory strength γ is tion history, i.e., number of interruption (NoI) and percentage
set to be 0.71. To show the impacts of a and b, we directly of interruption (PoI) that is defined as the interruption time
present the instantaneous playback quality qplay of different divided by total time.
levels in Table 1. Since the initial buffering can be regarded We show the evaluation results in Table 3. It can be
as a special case of playback stalling, the instantaneous play- seen that the proposed algorithm generally has better play-
back quality during initial buffering is calculated using (3) back quality than the reference algorithms. This is because
with constant Lini = −0.5. The initially selected quality the proposed algorithm attempts to enhance the QoE contin-
level are level 3 for all users. The playback threshold of buffer uum of all users by assigning reasonably higher quality video
size is 4 seconds. The simulation runs 200 seconds. to those users who suffered from the previously bad experi-
4.1. Playback Smoothness ence. Thus those users can quickly recover from the bad ex-
perience while keeping smooth playback, as shown in Fig. 2d.
We employ the number of quality level changes (NoC) and
Nevertheless, the aggressiveness in reference algorithms may
the playback smoothness (PS), a metric inherited from [7], to
result in interruption due to the estimation error of transmis-
evaluate the video consistency. PS is defined as the expected
sion rate. We present the buffer status of UE2 as an example
 of one playback round without level change, i.e., PS =
length
in Fig. 3. The proposed algorithm can quickly respond to
P 2
p=1 (np )/P . Here the continuous playback of one level the channel variation. For example, at around the 160th sec-
is defined as one round and it consists of np frames. There are ond, the channel status of UE2 is suddenly becoming bad (as
P rounds in total. Level 0 represents the playback stalling. shown in Fig. 2a). The proposed scheme can appropriately
From Table 2, we can see that the proposed algorithm choose the level to match such channel variation since we use
demonstrates the least NoC and highest PS. Besides, both a fine-grained representation of the video files. However, the
users enjoy smooth playback. However, for the reference reference algorithms fail to respond to this change and finally
 CQI variation of the celluar network Video level variation of Baseline algorithm Video level variation of InstRate algorithm Video level variation of the proposed algorithm
Average CQI during one segment

25 6 6 6
UE1 UE1 UE1
UE2 UE2 UE2
20 5 5 5

Video level index

Video level index

Video level index

15 4 4 4

10 3 3 3

5 2 2 2
UE1
UE2
0 1 1 1
0 10 20 30 40 50 60 70 80 90 100 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
Segment index Time (s) Time (s) Time (s)

(a) (b) (c) (d)

Fig. 2. (a) Channel CQI versus segment index; (b-d) The video quality level variation versus time.
Video data in the playback buffer of UE2 6. REFERENCES
250 Proposed
InstRate [1] R. Mok, X. Luo, E. Chan, and R. Chang, “Qdash: a
Video data size (KByte)

Baseline
200 qoe-aware dash system,” in Proc. of ACM MMSys, Feb.
2012, pp. 11–22.
150 [2] S. Akhshabi, A. Begen, and C. Dovrolis, “An experi-
100
mental evaluation of rate-adaptation algorithms in adap-
tive streaming over http,” in Proc. of ACM MMSys, Feb.
50 2011, pp. 157–168.
[3] O. Oyman and S. Singh, “Quality of experience for http
0 adaptive streaming services,” IEEE Commun. Mag., vol.
0 20 40 60 80 100 120 140 160 180 200
Time (s) 50, pp. 20–27, Apr. 2012.
[4] R. Houdaille and S. Gouache, “Shaping http adaptive
Fig. 3. UE2’s buffer size versus time
streams for a better user experience,” in Proc. of ACM
MMSys, Feb. 2012, pp. 1–9.
Table 4. Normalized Difference Between Users [5] C. Liu, I. Bouazizi, and M. Gabbouj, “Segment duration
UE1 and UE2
Metrics for rate adaptation of adaptive http streaming,” in Proc.
Baseline InstRate Proposed of IEEE ICME, July 2011, pp. 1–4.
APQ 6.14% 8.24% 2.04% [6] R. Kuschnig, I. Kofler, and H. Hellwagner, “Evalua-
PS 40.48% 30.44% 5.22% tion of http-based request-response streams for internet
video streaming,” in Proc. of ACM MMSys, Feb. 2011,
undergo stalling at around the 165th second. pp. 245–256.
4.3. QoE fairness [7] S. Xiang, L. Cai, and J. Pan, “Adaptive scalable video
We have implicitly shown the QoE fairness in the results pre- streaming in wireless networks,” in Proc. of ACM MM-
Sys, Feb. 2012, pp. 167–172.
sented earlier. Now we evaluate the QoE fairness using the
[8] A.El Essaili, D. Schroeder, D. Staehle, M. Shehada,
normalized difference of a certain metric between the two
W. Kellerer, and E. Steinbach, “Quality-of-experience
users, i.e., the difference of two values divided by the larger driven adaptive http media delivery,” in Proc. of IEEE
value. The normalized difference of APQ and PS is shown in ICC, June 2013, pp. 2480–2485.
Table 4. We observe that the proposed algorithm shows the [9] W. Pu, Z. Zou, and C. W. Chen, “Video adaptation proxy
least QoE difference between two users and thus guarantee for wireless dynamic adaptive streaming over http,” in
the QoE fairness. This is because higher priority is given to Packet Video Workshop, May 2012, pp. 167–172.
users with previously bad experience and that resources can [10] D. Rodriguez et al., “Quality metric to assess video
be saved from one user when the variation penalty is adopted. streaming service over tcp considering temporal loca-
tion of pauses,” IEEE Trans. Consum. Electron., vol.
5. CONCLUSION 58, pp. 985–992, Aug. 2012.
In this paper, we propose a QoE continuum driven quality [11] ITU, “ITU-R BT.500-13 methodology for the subjective
adaptation algorithm to overcome the challenges resulting assessment of the quality of television pictures,” 2012.
from imprecise QoE monitoring, unfair QoE, and inaccurate [12] J. Xue, D. Zhang, H. Yu, and C. W. Chen, “Assessing
video representation in multi-client wireless HAS. By em- quality of experience for adaptive http video stream,” in
ploying the unaggressive, fair, and fast-responding adaptation Proc. of IEEE ICME workshop on emerging multimedia
systems and applications, accepted, 2014.
logic, the proposed algorithm outperforms existing works and
[13] 3GPP, “3GPP TS 25.214 v7.1.0 physical layer proce-
achieves satisfactory QoE and fairness. Future work shall be dures(fdd),” 2006.
focused on extending the algorithm to larger-scale systems, [14] Enhanced UMTS radio access network extensions for ns
wherein downlink scheduling can also be incorporated. Be- 2-User Guide (Release 1.6), [Online] Available:
sides, the framework parameters need to be optimized in order http://eurane.ti-wmc.nl/eurane.
to further enhance the overall system performance.