In-Network Quality Optimization for
Adaptive Video Streaming Services
Niels Bouten, Student Member, IEEE, Steven Latr
´
e, Member, IEEE, Jeroen Famaey, Member, IEEE,
Werner Van Leekwijck, and Filip De Turck, Senior Member, IEEE
Abstract—HTTP Adaptive Streaming (HAS) services allow the
quality of streaming video to be automatically adapted by the
client application in face of network and device dynamics. Due
to their advantages compared to traditional techniques, HAS-
based protocols are widely used for Over-The-Top (OTT) video
streaming. However, they are yet to be adopted in managed envi-
ronments, such as ISP networks. A major obstacle is the purely
client-driven design of current HAS approaches, which leads
to excessive quality oscillations, suboptimal behavior, and the
inability to enforce management policies. Moreover, the provider
has no control over the quality that is provided, which is essential
when offering a managed service. This article tackles these chal-
lenges and facilitate the adoption of HAS in managed networks.
Specifically, several centralized and distributed algorithms and
heuristics are proposed that allow nodes inside the network to
steer the HAS client’s quality selection process. The algorithms
are able to enforce management policies by limiting the set of
available qualities for specific clients. Additionally, simulation
results show that by coordinating the quality selection process
across multiple clients, the proposed algorithms significantly
reduce quality oscillations with a factor 5 and increase the average
delivered video quality with at least 14%.
I. INTRODUCTION
T
HE increasing popularity of Over-The-Top (OTT) mul-
timedia services has led to the widespread adoption of
HTTP-based streaming protocols. Such protocols have many
advantages compared to traditional streaming methods, such
as reuse of existing HTTP infrastructure (e.g., servers, proxies
and caches), reliable transmission and firewall compatibility.
Originally, progressive download techniques were used, al-
lowing the user to start consuming the content after a few
seconds of buffering. However, progressive download methods
cannot cope with congestion, the highly fluctuating throughput
of mobile networks or diverging characteristics of devices
and networks. To overcome said problems, a new generation
of HTTP-based streaming protocols, collectively referred to
as HTTP Adaptive Streaming (HAS), was introduced. The
offered content is split into a set of temporal segments, which
are encoded at multiple bit rates. In traditional HAS, a rate
Copyright (c) 2013 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to pubs-permissions@ieee.org.
N. Bouten and F. De Turck are with the Department of Electrical and Computer
Engineering, Ghent University - iMinds, Belgium. S. Latr
´
e and J. Famaey are
with the Department of Mathematics and Computer Science, University of
Antwerp - iMinds, Belgium. W. Van Leekwijck is with Alcatel-Lucent Bell
Labs, Antwerp, Belgium
E-mail: niels.bouten@intec.ugent.be
adaptation algorithm, deployed at the client, is then used to
select the bit rate of each segment, based on the current
network conditions, buffer status and device capabilities.
State-of-the-art HAS solutions embed the rate adaptation
algorithm inside the client application. This allows the client
to independently choose its playback quality and prevents the
need for intelligent components inside the network, which are
the main reasons HAS is used in OTT scenarios. However,
academia and industry are showing a growing interest in the
use of HAS in managed networks [1]
12
, for example by opti-
mizing the delivery by applying in-network bitrate adaptation
3
or by deploying IP multicasting to ease the distribution of
linear TV HAS services [2]
4
. The extensive content catalogue
and increased flexibility in terms of supported devices of
these OTT-services (e.g., YouTube, Hulu, Netflix) but deliv-
ered over the managed network, could greatly benefit both
the provider and the end-user. However, in such environ-
ments, a purely client-driven approach has several significant
disadvantages. First, the lack of coordination among clients
leads to competing behavior among those clients, resulting
in incorrect throughput estimations, causing excessive qual-
ity oscillations and suboptimal decisions [3], [4], negatively
impacting QoE [5]. Second, management policies, such as
user subscription constraints and guarantees on the delivered
quality, cannot be easily enforced [6], [7]. In order to facilitate
adoption of HAS for the delivery of multimedia services in a
managed environment, these challenges should be tackled.
A straightforward solution to the resource scarcity affecting
streaming services could be to increase the physical capacity
of the delivery network. These updates are however associated
with high costs for the service provider, while an in-network
optimization based solution does not affect these infrastructure
costs. Since technologies (e.g. the advent of Ultra High Defi-
nition Television streaming) are constantly evolving, frequent
infrastructure updates are required to cope with the ever
increasing traffic demands. Physical infrastructure upgrades are
time-consuming, therefore we believe that there should be a
coexistence of both approaches to deal with future demand
by intelligently managing resources in attendance of physical
capacity updates.
This article proposes a hybrid approach where the rate
1
http://www.juniper.net/us/en/local/pdf/solutionbriefs/3510463-en.pdf
2
http://www.rgbnetworks.com/pdfs/RGB-Velocix Adaptive Streaming
CDN White Paper 0911-01.pdf
3
http://www.cachelogic.com/vx-portfolio/solutions/velocixeve
4
http://www.velocix.com/vx-portfolio/solutions/video-optimised-
architecture
adaptation algorithm is steered by an in-network component
to address the aforementioned challenges. It is deployed on
intermediary proxies and supports client-side rate adaptation
algorithms by dynamically limiting the possible set of bit rates
to select from. Currently an operator’s multimedia delivery
network typically contains several transparent caches and QoE
measurement platforms which interpret HTTP headers and
reconstruct HTTP adaptive streaming sessions in order to
evaluate the end-to-end QoE. These platforms can be extended
to not only measure the QoE, but also optimize the QoE
by performing in-network quality optimization, thus requiring
only limited extensions to the already available infrastructure.
The proposed hybrid approach allows clients to still react
upon sudden network changes or scarcity in device resources,
while increasing the overall quality and stability. Moreover, it
can enforce a wide range of management policies, allowing
providers to specify priorities when allocating resources to a
certain group of users. This translates into several concrete
contributions. First, the in-network rate adaptation problem is
formally defined. Second, an optimal centralized algorithm is
proposed that solves the problem as an Integer Linear Program
(ILP). Third, a scalable variant of the algorithm is introduced
that can be distributed across multiple logically hierarchical
intermediary proxies. Finally, a heuristic with significantly
lower computational complexity is proposed.
The remainder of this article is structured as follows. Sec-
tion II lists and discusses state of the art research on client-
based and in-network HAS rate adaptation. Subsequently, the
in-network rate adaptation problem is formally defined in
Section III. Sections IV and V describe and evaluate the three
algorithms proposed to solve this problem respectively. Finally,
Section VI concludes the article.
II. RELATED WORK
The increased popularity of video consumption over the
Internet has led to the development of a range of protocols
that allow adaptive video streaming over HTTP. Some of the
major players have introduced their proprietary protocols such
as Microsoft’s Silverlight Smooth Streaming
5
, Apple’s HTTP
Live Streaming
6
and Adobe’s HTTP Dynamic Streaming
7
.
More recently, a standardized solution has been proposed
by MPEG, called Dynamic Adaptive Streaming over HTTP
(DASH) [8]. Although differences exist between these imple-
mentations they are based on the same basic principles: a
video is split up into temporal segments which are encoded
at different quality rates, the autonomic video client heuristic
then dynamically adapts the quality, based on metrics such
as average throughput, delay and jitter. The drawback of this
approach is of course that all control lays in the hands of
the clients which strive to maximize their individual quality.
From the providers perspective however, other factors such as
5
Microsoft Smooth Streaming - http://www.iis.net/downloads/microsoft/
smooth-streaming
6
Apple HTTP Live Streaming - http://tools.ietf.org/html/
draft-pantos-http-live-streaming-13
7
Adobe HTTP Dynamic Streaming - http://www.adobe.com/products/hds-
dynamic-streaming.html
minimization of costs and prioritization of users with higher
subscription levels are of equal importance. Current HAS
approaches do not support intervention during the quality
assignment process which is fully dominated by the clients.
The approach presented in this paper therefore focuses on
managing the quality for HAS by the service provider.
The performance of HAS-based services can be improved
by applying changes both at the client and the delivery
network. Each commercial HAS implementation comes with
a proprietary client heuristic. Akhshabi et al. compare sev-
eral commercial and open source HAS players and indicate
significant inefficiencies in each of them, such as frequent
oscillations and unfairness when the number of competing
clients increases [3]. Several heuristics have been proposed
in literature as well, each focussing on a specific deployment.
Liu et al. discuss a video client heuristic that is suited for a
Content Delivery Network (CDN) by comparing the expected
segment fetch time with the experienced segment fetch time to
ensure a response to bandwidth fluctuations in the network [9].
Andelin et al. provide a heuristic which was specifically
designed for Scalable Video Coding (SVC) and using a slope
to define the trade-off between downloading the next segment
and upgrading a previously downloaded segment [10]. In
previous work [11][12], the authors evaluated different client
heuristics both for Advanced Video Coding (AVC) and SVC,
applying optimizations such as pipelined and parallel download
scheduling. Several of the aforementioned authors indicate the
impact of competing HAS clients on the quality oscillations,
which are known to have a negative impact on Quality of
Experience (QoE) [5]. Furthermore, most of the commercial
client heuristics require a considerably large buffer to be able
to react to network changes. This paper therefore aims at
controlling the quality by introducing global QoE management,
reducing drastically the number of quality oscillations and
allowing to reduce the required buffer size. The presented
approach is applicable to both AVC and SVC.
An autonomic delivery framework for HAS-based Live TV
and Time Shifted TV (TSTV) was presented in previous
work [13][2] which allows to reduce the consumed bandwidth
by grouping unicast HAS sessions sharing the same content
into a single multicast session. However, for Video on Demand
(VoD) HAS sessions, the content is more diverse and only
few sessions are potentially shared among multiple users. This
prevents them to be grouped into a shared multicast session
and therefore prevents them from being delivered in a scalable
manner. In [14], an overview of interesting use cases for
applying SVC in a network environment are presented, among
which the graceful degradation of videos when the network
load increases. The authors argue the need for Media Aware
Network Elements (MANEs), capable of adjusting the SVC
stream based on a set of policies specified by the network
provider. Similar to this approach, Latr
´
e et al. proposes an in-
network rate adaptation algorithm, responsible for determining
which SVC quality layers should be dropped in combination
with a Pre-Congestion Notification (PCN) based admission
control mechanism [15]. In [16], a prototype of an intermediary
adaptation node is proposed, where the media gateway esti-
mates the available bandwidth on the client link and extracts
the supported SVC-streams. Similar to this, the WiDASH
proxy is responsible for in-network video adaptation and is
able to perform global optimization over multiple concurrent
HAS flows by prioritizing clients which have poor channel
quality [17]. Wirth et al. discuss the optimization of multi-
user resource assignment for DASH video transmission over
the LTE downlink [18]. By deploying a cross-layer technique
for allocating the resources at the base station and taking into
account the specific information of the video sessions, the
number of playout starvations can be considerably reduced.
In Parakh et al., the authors propose a game theoretic approach
towards decentralized bandwidth allocation for video streams
in wireless systems, where users are charged for bandwidth re-
sources proportionally to the requested bit-rate [19]. Situnen et
al. propose dropping video layers based on their priority when
network congestion arises for scalable video streaming over
wireless networks [20]. Most of the aforementioned research
focuses on the dropping of quality layers when congestion
arises, meaning the quality is limited in the same way for all
users. Our proposed approach limits the maximum quality in a
per client manner, allowing the service provider to differentiate
the delivered video services based on the clients subscription.
This allows the service provider to control the QoE on a per
subscriber level, and thus offering different subscription types
for the VoD HAS services.
Lee et al. describe a three-tier streaming service where
multiple clients are connected through multiple intermediate
proxies to a multimedia server [21]. The authors only consider
live streaming, if however VoD streaming would be targeted,
the streaming service can no longer be delivered in an efficient
way using multicast streaming, since a lot of requests are on
unpopular content which is infrequently requested. This causes
the content to be delivered using unicast transport from origin
to regional servers and thus having the risk of running into
bandwidth bottlenecks on these links as well, which is not
addressed within the cited paper. Furthermore, videos need to
be transcoded in the intermediary proxies, in standard HAS
however, the quality levels are discrete and fixed, causing
the objective function in the proposed solution to change
drastically and leading to the inability to use the max-min
composition. Unlike Real Time Streaming Protocol (RTSP)
en Real Time Transport Protocol (RTP) based streaming pro-
tocols, there is no server-side bitrate adaptation required, the
client decides autonomously which quality it will select, based
on the current state of the network and from a list of permitted
qualities, selected from within the network. This also implies
that if a client struggles to achieve its assigned quality level,
for example due to poor wireless connection quality or limited
CPU resources, it can still decide to switch to a lower quality.
In [22][23], the authors focus on optimizing the allocation
of bits of video sequences among multiple senders to stream
to a single client. Peer-to-peer streaming and multi-server
distributed streaming are the main use-cases of this approach,
there is no simple extension of the work when multiple
clients need to share the same server side bottleneck. Fur-
thermore, this requires fine-grained scalable video streaming
to support the allocation of non-overlapping bit ranges to
multiple servers, while for HAS, fixed bitrate representations
…"
…"
…"
…"
…"
Legend"
Fig. 1: Graphical representation of variables and assumptions.
are available, encoded using advanced video coding, leading to
video segments of which the quality cannot be improved in a
straightforward way by downloading additional bit ranges. Our
work however, could also be extended to support scalable video
in a straightforward way. Akhshabi et al. propose server-side
rate adaptation to cope with unstable streaming players due to
ON-OFF patterns when they compete for bandwidth [24]. The
systems detect sudden rate fluctuations in the client playout
and try to solve them by shaping the sending rate at the
server to resemble the bitrate of the stream. These systems
are able to restore the streaming session when oscillation
or freezing occurs and then remove the shaping when the
client has stabilized. Our approach is not only able to solve
the problems of oscillation or freezes when they occur, but
actively tries to prevent them. This is because we can use more
detailed in-network information. This article is an extension
to our previous work on in-network quality management for
HAS [25]. However, the problem formulation is generalized
and we significantly extended the approach with a centralized,
distributed and relaxed optimization. Furthermore, the previous
work only considered simple topologies with a single bot-
tleneck where multiple clients directly connect to the server.
Whereas this article supports more complex topologies with
multiple levels, multiple bottlenecks and intermediary proxies,
as well as asymmetric topologies.
III. FORMAL PROBLEM DESCRIPTION
Providers are exploring how they can offer VoD HAS
services next to traditional TV services over their managed
network environment. HAS services offer the same content
at multiple qualities, each at their corresponding rate. This
allows providers to perform QoE management by adjusting
each sessions quality level, based on the current network
utilization. At peak times, the consequences of an inadequate
amount of resources in the network, can thus be anticipated
by reducing the quality of individual streaming sessions, while
still allowing admittance of all users.
TABLE I: Variables used for the rate decision
α
s
Weighing factor to model the tradeoff between quality and quality
switches
B
e
Bandwidth reserved for HAS traffic for edge e ∈ E
β
q
Bitrate associated with layer q ∈ Q
β
max
Highest bitrate in Q
C ⊂ N The set of HAS VoD clients
C
e
⊆ C The set of clients in the service delivery tree for which the VoD
traffic traverses edge e ∈ E
C
n
⊆ C The set of clients in the service delivery tree for which the VoD
traffic traverses node n ∈ N
E The set of edges in the service delivery tree
E
c
⊆ E The unique delivery path from server s
c
to client c ∈ C
e
n
−
The edge connecting node n to its predecessor n
−
E
n
+
⊆ E The set of edges connecting node n to its successors N
n
+
H
c
The history of previous quality decisions for client c ∈ C
h
c,q,t
∈
H
c
Binary variable indicating wether client c ∈ C was assigned quality
q ∈ Q at time t
N The set of nodes in the service delivery topology
N
n
+
⊂ N The set of successors of node n
n
−
The predecessor of node n
P ⊂ N The set of proxies in the delivery tree
Q Available quality rates for video
Q
v
⊆ Q Available quality rates for video v ∈ V
S ⊂ N The VoD access server
s
c
∈ S The VoD access server for client c
V The set of available videos via VoD server S
v
c
∈ V The video v ∈ V for which client c is requesting access
A. Definition of variables and assumptions
Figure 1 gives an overview of the problem variables and
assumptions. Let us consider an access network topology
modeled as a graph, consisting of a set of nodes N , which
encompasses servers S ⊂ N , proxies P ⊂ N , and clients
C ⊂ N . A set of edges E connects the nodes in a logical tree
topology which is typically used for video delivery networks
8
.
Note that typical access networks are using a logical tree for
their delivery, although the underlying physical network is not
a tree due to replication concerns. Every node n ∈ N has
an incoming edge e
n
−
∈ E connecting it to its predecessor
n
−
∈ N and a set of outgoing edges E
N
+
⊆ E connecting it
to its successors N
+
⊂ N . Every edge e ∈ E has an associated
bandwidth capacity B
e
reserved for HAS traffic.
The servers host a set of videos V. Every video v ∈ V has an
associated set of quality representations Q
v
⊆ Q. Moreover,
every quality representation q ∈ Q has a bit rate β
q
. Every
client c ∈ C has an associated origin server s
c
∈ S, a unique
delivery path E
c
⊆ E from that server, and a video v
c
∈ V. The
set of clients that have an edge e ∈ E as part of their delivery
path E
c
, is represented by C
e
⊆ C. In summary, Table I lists
the symbols introduced throughout this section.
B. Integer Linear Programming formulation
The problem consists of maximizing the QoE over all clients
c ∈ C, while adhering to the edge bandwidth constraints. The
solution is characterised by a boolean decision matrix A. The
element a
c,q
∈ A is equal to 1 if quality q ∈ Q
v
c
is selected
for client c ∈ C, and 0 otherwise. The decision variables are
8
An example is the Triple Play Service Delivery Architecture from Alcatel-
Lucent (http://goo.gl/4aZVvf), which is used by over 50 operators worldwide
(http://goo.gl/kHMY1b)
subject to the following two constraints:
∀c ∈ C, ∀q ∈ Q
v
c
: a
c,q
∈ 0, 1 (1)
∀c ∈ C :
X
q∈Q
v
c
a
c,q
= 1 (2)
The above constraints state that the decision variables are
boolean values and that only one quality representation can
be selected per client.
According to Padhye et al., the maximum achievable
throughput B for a TCP connection subject to a round trip time
RT T and maximum window size W
max
, probability of packet
loss p, delayed ACK number of b and average retransmission
timeout T
0
can be approximated by the following [26]:
B(p) ≈ min
W
max
RT T
,
1
RT T
q
2bp
3
+ T
0
min
1, 3
q
3bp
8
p(1 + 32p
2
)
(3)
The maximum achievable TCP throughput for client c is
thus limited by its window size W
max,c
and its RT T
c
. Both
parameters can be estimated or measured at the client and
forwarded to the in-network control proxy. When p values
are low, which is the case in fixed networks, the achievable
throughput is primarily limited by the first term. To limit
the overhead of acquiring packet loss probabilities for all
clients, only the first part of the TCP estimator is considered.
Therefore, the following constraint is added, limiting the end-
to-end achievable throughput for each client:
∀c ∈ C :
X
q∈Q
c
a
c,q
× β
q
≤
W
max,c
RT T
c
(4)
When N TCP-connections use the same bottleneck link,
Altman et al. state that the maximum aggregated achievable
throughput that can be obtained is a factor of the link capacity
B
e
[27]:
B
max
≈
1 −
1
1 + cN
B
e
(5)
Where c =
1+d
1−d
with d the fraction with which the
send rate is decreased when congestion arises. We use this
approximation of the maximum achievable throughput to limit
the aggregated allocated rate of the different clients:
∀e ∈ E :
X
c∈C
e
X
q∈Q
c
a
c,q
× β
q
≤
1 −
1
1 + c |C|
B
e
(6)
As stated, the objective aims to maximize the global QoE.
This is obviously a broad term that can be interpreted in
a multitude of ways. As such, a generic objective function
is proposed that can be adapted to the service provider’s
optimization policy, represented by the function F (·):
max
X
c∈C
X
q∈Q
v
c
F (a
c,q
) (7)
For example, the provider could aim to maximize the total
delivered bit rate, which can be translated into the following
objective function:
max
X
c∈C
X
q∈Q
v
c
a
c,q
× β
q
(8)
The operator could also decide to optimize the fairness
among the connected clients. This can be achieved by adopting
proportional fairness, as proposed by Kelly et al. [28][29].
A vector of rates A
c
= (
P
q∈Q
v
c
a
c,q
× β
q
, c ∈ C) is
proportionally fair if it is feasible, according to Equation (4)
and (6) respectively, and if for any other feasible vector A
∗
c
,
the aggregate of the proportional changes is zero or negative:
X
c∈C
A
∗
c
− A
c
A
c
≤ 0 (9)
According to Wei et al., the fair bandwidth allocation can
be represented by a local maximum of the logarithmic utility
function [30]. Since this function is differentiable and strictly
concave, it has only one maximum, which is therefore also the
global maximum. Therefore, the objective of a proportionally
fair bandwidth allocation can be expressed by:
max
X
c∈C
log (
X
q∈Q
v
c
a
c,q
× β
q
) (10)
Another objective could be to minimize the number of
switches since they have a negative impact on overall QoE.
This requires maintaining a history H
c
of previous quality
decisions for each client c ∈ C where h
c,q,t
= a
c,q
at time
t. For quality switches, not only the frequency of switching
is important, but also the distance between quality selections
affects the overall quality [5]. Therefore, to assess the impact
of distance in quality, the variation in quality over the history
H
c
is taken into account. The following weighted sum is
used to model the impact on switching behavior, where µ
represents the average quality, σ introduces a penalty for
quality switching and α
s
represents a weighing factor used
to emphasize either the impact of quality or the switching
behavior:
max α
s
× µ − (1 − α
s
) × σ (11)
Since the decision variables a
c,q
are binary variables, the
calculation of the objective function can be simplified by
calculating µ
c,q
and σ
c,q
for each client c and it’s associated
quality range Q
c
. The quality rates are normalized with respect
to the highest quality rate β
max
.
∀c ∈ C, ∀q ∈ Q
v
: µ
c,q
=
1
|H
c
| + 1
β
q
β
max
+
X
h
c,t
∈H
c
X
q∈Q
c
h
c,q,t
× β
q
β
max
(12)
In this way, the use of quadratic terms in the objective
function is avoided. The penalty σ for switching between
qualities can be calculated as follows:
∀c ∈ C, ∀q ∈ Q
v
: σ
c,q
=
s
1
|H
c
| + 1
∗
v
u
u
t
q × β
q
β
max
− µ
c,q
2
+
X
h
c,q,t
∈H
c
X
q∈Q
c
h
c,q,t
× β
q
β
max
− µ
c,q
2
(13)
The total objective can then be expressed as:
max
X
c∈C
X
q∈Q
v
a
c,q
× (α × µ
c,q
+ (1 − α) × σ
c,q
) (14)
IV. ALGORITHMS
A. Centralized ILP
The ILP formulation described in Section III-B can be
used to optimize the quality assignments using a centralized
controller. It requires as input the knowledge of the delivery
network topology (N , E), link constraints B
e
, the set of clients
C
e
for which the VoD traffic traverses an edge e ∈ E and the
characteristics of these clients (W
max,c
, RT T
c
). Solving said
ILP formulation will yield a set of optimal quality assignments
a
c,q
for each client c and quality level q. These assignments
are optimal in the sense that they maximize the objective
P
c∈C
P
q∈Q
F (a
c,q
) subject to the constraints described in
Equations (1), (2), (4) and (6). As we assume there is a
constant bitrate reserved for HAS traffic on each edge, the
Centralized optimization is executed each time a newly joined
client requests a manifest file or if a client becomes inactive
by leaving the delivery network.
B. Distributed ILP
The number of constraints for the centralized ILP grows
with an increasing depth of the service delivery topology tree.
Consider a topology tree with k child nodes per node and
l levels (thus l = log
k
|C| + 1), the total number of edge
constraints is then equal to
P
l−1
i=0
k
i
which can be written as
1−k
l
1−k
. This leads to an exponentially increasing model size
with the number of levels in the delivery tree, affecting the
calculation time. Since our approach would be deployed in
an operational setting, the decision process should be able
to determine quality allocations in real-time. Therefore, we
propose a distributed approach, where each proxy locally
determines the optimal allocation constrained by the local edge