A BIO-INSPIRED HTTP-BASED ADAPTIVE STREAMING PLAYER
Yusuf Sani
∗
, Andreas Mauthe, Christopher Edwards
School of Computing and Communications
InfoLab2l, Lancaster University
Lancaster LA1 4WA, UK
y.sani,a.mauthe,c.edwards @lancaster.ac.uk
Mu Mu
Faculty of Arts, Science and Technology
The University of Northampton
Northampton NN2 6JB, UK
mu.mu@northampton.ac.uk
ABSTRACT
In order to streamline video content distribution on a myr-
iad of platforms over heterogeneous networks, HTTP Adap-
tive Streaming (HAS) is being increasingly adopted. In this
paper we pilot a bio-inspired HAS optimisation design with
the aim of maximising the overall user experience of a video
playback session. Evaluations conducted within a real-world
Internet environment, using quality indicators such as conver-
gence time, start-up delay, average video rate, stability, and
fairness, demonstrate the benefits of our design.
Index Terms— HTTP Adaptive Streaming, Adaptive Bi-
trate Selection, Buffer-based Player
1. INTRODUCTION
HTTP-based adaptive video streaming service ensures that
the quality of video each user receives is derived from con-
text dependant information, e.g. throughput, buffer level, etc.
This requires a client to continuously monitor and estimate
the resource availability. It then requests a chunk with the
highest video rate the estimated network and device capac-
ity can sustain. This unique mechanism allows HAS video
services to adapt to different environments and changes in
transmission channel states. The main challenge of design-
ing an adaptation scheme is to seek an optimal balance be-
tween a number of QoE factors such as video rate and sta-
bility or occurrence of re-buffering. Inspired by the study of
population dynamics in biology, this paper introduces a novel
HAS player that maximises user experience. Evaluations con-
ducted within a real-world Internet environment demonstrate
the benefits of our design with respect to fast convergence,
low start-up delay, maximum utilisation, high average video
rate, high stability, fairness to other players and video freeze
avoidance.
2. BACKGROUND
HAS services provide audio-visual content in one or multi-
ple adaptation sets of representations. Each representation is
∗
Thanks the Petroleum Technology Development Fund (PTDF), Nigeria
for funding.
encoded using a specific encoding configuration (such as bi-
trate and native frame resolution). A representation of content
is stored in a series of chunks (segments). Using a manifest
file such as Media Presentation Description (MPD), a HAS
client progressively requests chunk of a certain bitrate, from
the selected representation, based on the level of available re-
sources measured at the client side. Such a streaming method
is also known as adaptive bitrate selection (ABR). The first
generation of ABRs select a chunk with the bitrate just below
the estimated throughput derived from a predefined number of
previously downloaded chunks [1, 2]. When there is a change
of the estimated throughput, the media playback buffer level
can be used as an indicator to determine whether to increase,
decrease or keep the current quality level in order to max-
imise the playback quality whilst avoiding buffer under-run
(i.e., re-buffering) [3, 4]. Many of the current HAS services
are found to suffer from unnecessary re-buffering [5], unde-
sirable fluctuations [6] and sub-optimal video quality [6]. In-
stead of solely relying on heuristic, it is shown in [7] that
incorporating a system model that captures the relationship
between buffer state changes and video rate can improve the
QoE performance of a HAS player. Based on this observa-
tion we were motivated to take a cross-disciplinary approach
towards the QoE modelling of ABR algorithms.
Population dynamics is a branch of mathematical ecology
that is mainly concerned with how the population of species
grows and shrinks as well as the processes that influence the
changes e.g. resource and competing specie [8]. This pa-
per focuses on the most well-known population model, the
Verhulst-Pearl logistic model. The model is very well under-
stood and is known to be stable and converges at the maxi-
mum carrying capacity. We then propose a buffer-based ABR
algorithm that uses the Verhulst-Pearl equation to drive its rate
evolution map.
3. VERHULST-PEARL BASED VIDEO RATE MAP
The video quality level is assumed to be a specie whose
growth we are interested in, and the playback buffer is con-
sidered to be its habitat. We assume that the rate at which the
incoming video chunks arrives is analogous to birth rate and
the rate at which a player consumes content from the play-
Fig. 1. Experimental Set-up
back buffer is assumed to be the death rate. It is known that
the rate of chunk arrival is
c(t
i
)
Q
[5], where c(t
i
) is the through-
put of a link and Q is bitrate of the chunk being downloaded.
Please note that the consumption rate is constant at one sec-
ond of content every wall-clock second. Therefore, buffer
level changes B
t
at any time t depends on these two param-
eter. AS in a natural habitat, the buffer size will determine
a limit on the maximum video rate q
max
a player can down-
load. In our case, unlike in the wild, the maximum video rate
is given (as defined in the MPD), therefore our task is focused
on identifying the amount of the buffer required to guarantee
the maximum video rate. We now modify the Verhulst-Pearl
equation thus as:
dQ
dB
= αq(q
max
− q) (1)
Equation (1) defines the only valid system state-space,
which is represented by the rate at which video rate changes
with respect to buffer state. With α being the growth con-
stant and q the requested video rate. To derive the rate map
we solve the differential equation (1) assuming that, like in a
natural context, there is a seed population that reproduces to
kick-start the growth. We call this the minimum video quality
level q
0
.
Q =
q
max
1 + [
q
max
q
0
− 1]e
−αq
max
B
t
(2)
It is easy to show that equation (2) will always converge
at the maximum video rate, that is, lim
B→∞
Q = q
max
[7].
3.1. Implementation
A HAS player chooses video representations from a finite and
discrete set. We first restrict the player to switch between ad-
jacent video rates only so as to prevent high amplitude varia-
tions.
Assuming q
k
has been completely downloaded (please re-
call that we are starting with (q
min
), we can use equation (3)
to calculate the buffer distance needed to change video rate.
Once ∆B
k+1
= 0 we move up the ladder of video rates. If
∆B
k−1
= 0, the video rate is switched down, else we main-
tain the current video rate.
∆B
k±1
=
Q
−1
(q
k+1
) − B
t
q
k+1
B
t
− Q
−1
(q
k−1
) q
k−1
(3)
4. PERFORMANCE EVALUATION
The experimental set-up is shown in Figure 1. The client is
connected to the Internet either via an Ethernet switch or EE’s
3G networks in the UK. We use the web services located at
the Alpen-Adria-Universit
¨
at, Klagenfurt, which hosts the Big
Buck Bunny dataset [9]. The player is implemented in Python
and uses Request package for HTTP request-response trans-
actions, and is tested on Ubuntu 12.04.2 LTS with 3.8 kernel.
The host that runs the player also hosts dummynet, tcpdump,
lsof, and wget as part of the experiment.
We conducted a “blue-sky” test in both wired and wireless
setting. Then varies the link capacity using dummynet. We
set B
max
= 200s, and the the growth constant α = 0.05
is used. Each experiment was conducted 10 times and the
average result is used. When more than one player is used
or when a player operates in the presence of a background
traffic, the experiment is conducted on the same machine for
the maximum portability.
4.1. Evaluation Metrics
There is a growing number of research in the field of QoE,
and no definite model has so far been established for adaptive
streaming [10, 11]. To evalaute the performance of our model,
we employ the following metrics which are widely recognised
as the key QoE indicators:
• Re-buffers: this is the total number of video freeze
event per streaming session.
• Average video rate: is calculated as
t
1
q
1
+t
2
q
2
...t
n
q
n
t
n
−t
1
and
measured in kb/s [12].
• Stability: measures the degree of video quality oscilla-
tion as described in [13].
• Utilisation of available network resource: is calculated
by dividing the average video rate by the average net-
work capacity [14].
• Fairness: two or more services are said to be fair if
they divide the available resource among themselves
equally.
• Convergence time: is the time taken to settle at the sus-
tainable video rate.
• Startup Delay: in this paper we define startup delay as
the amount of time it takes for a player to download 15
chunks before the playback starts.
5. RESULTS
This section discusses the result of the various test-bed exper-
iments conducted.
5.1. Startup Period and Stability
In the first set of the experimentation we stream in both the
wired and wireless environments with the network link as is.
As can be see from Figure 2 the player is able to converge
at the maximum video rate in a very short time and without
a single instance of oscillation within the wired environment.
Furthermore, the start-up delay is minimal at 1.19s. A similar
performance is obtained when the player is run in a wireless
environment (see Table 1 the summary). It’s worth noting
that the player seems to be more cautious in a risky environ-
ment, delaying the video rate upgrade in order to maximises
Table 1. Adaptation for Variable Bandwidth
Environment Startup
Delay
(s)
Convergence
Time (s)
Maximum
Rate
(kb/s)
Wired 1.19 16 8000
Wireless 1.25 91 8000
Fig. 2. Blue Sky Test (Wired Network)
stability. As can be seen from Figure 4(b) there are only three
(3) instances where the player changes its video rate, with the
oscillation factor of only 2.0%, even though the network is
anything but stable as can be observed from Figure 4(a).
5.2. Responsiveness and utilisation
Next, we investigate how the propose player adapts in a net-
work with severe bandwidth fluctuation. Figure 3 clearly
demonstrates that the player is able to converge at exactly
the system capacity regardless of whether it is an upward or
downward convergence. Furthermore, when the link capacity
suddenly decreases, the player does not instantly over-react
like many other HAS player designs, rather because it senses
the buffer can sustain high video rate it takes a modest ap-
proach, and when it starts reducing the video rate it does that
linearly. As is summarised in Table 2, the design maximises
the utilisation of link capacity at 106% without causing any
video freeze or resulting in buffer dropping to a risky level.
This confirms the fact that with careful design, a player can
download above the link capacity, provided that some content
is pre-buffered, without affecting its QoE performance.
5.3. Fairness
Also investigated is how fair the player is to other HAS player
and background TCP traffic. A total of four players, all using
the same implementation, were run concurrently. We set the
maximum bandwidth to 6Mb/s, the hypothesis is that if the
players are fair to one another they should equally share the
available bandwidth (because all the players are connected to
the same network and are also running on similar device). As
can be seen from Figure 5(a) none of the players secures more
than 1.5Mb/s, a fair share of the available capacity. This is
Fig. 3. Responsive and Utilisation test
Table 2. Adaptation for Variable Bandwidth
Rebuffers Maximum
Video rate
(kb/s)
Average
Video
rate(kb/s)
Throughput
Utilisation
(%)
0 6000 3827 106
achieved with a high level of stability, as none of the players
observed more than three unnecessary oscillation of the video
rate.
Next, the background traffic (file downloading from the
same server) is started 30s after the start of the streaming ses-
sion. As can be seen in Figure 5(b), as soon as the background
traffic is started the video quality starts to gradually drop un-
til an equilibrium is reached. They fairly share the available
bandwidth, that is, each uses about 3Mb/s. Furthermore, it
is worth noting that the drop in video rate does not affect the
stability of players, start-up delay or other QoE factors.
6. CONCLUSIONS
HTTP adaptive streaming provides a great foundation for on-
line media streaming over heterogeneous networks. Recent
years have seen an increasing amount of efforts in the devel-
opment of adaptation algorithms with the shared objectives
of maximising the quality of user experiences. This paper pi-
lots a new approach that balances multiple key quality factors
using a bio-inspired mechanism. Evaluation results demon-
strate the effectiveness of the proposed design with respect to
start-up delay, convergence time, stability, network efficiency,
average video quality and fairness.
7. REFERENCES
[1] Saamer Akhshabi, Ali C. Begen, and Constantine
Dovrolis, “An experimental evaluation of rate-
adaptation algorithms in adaptive streaming over http,”
in Proc. of the MMSys, 2011, pp. 157–168.
[2] Truong Cong Thang, Quang-Dung Ho, Jung Won Kang,
and Anh T Pham, “Adaptive streaming of audiovisual
content using mpeg dash,” Consumer Electronics, IEEE
Transactions on, vol. 58, no. 1, pp. 78–85, 2012.
(a) Blue Sky Test (wireless)
(b) Player Oscillation
Fig. 4. Four players streaming at the same time.
(a) Four Players streaming at the same time.
0
1000
2000
3000
4000
5000
6000
7000
8000
0 50 100 150 200 250 300 350
0
20
40
60
80
100
120
[Kb/s]
Buffer Level (s)
Elapse Time(s)
Bitrate
Throughput
Bandwidth
Buffer Occupancy
(b) Player and Background TCP traffic.
Fig. 5. Fairness test
[3] Saamer Akhshabi, Lakshmi Anantakrishnan, Ali C.
Begen, and Constantine Dovrolis, “What happens
when http adaptive streaming players compete for band-
width?,” in Proc. of the NOSSDAV, 2012, pp. 9–14.
[4] Junchen Jiang, Vyas Sekar, and Hui Zhang, “Improving
fairness, efficiency, and stability in http-based adaptive
video streaming with festive,” in Proc of the 8th Inter-
national Conference on Emerging Networking Experi-
ments and Technologies. ACM, 2012, pp. 97–108.
[5] Te-Yuan Huang, Ramesh Johari, and Nick McKeown,
“Downton abbey without the hiccups: Buffer-based rate
adaptation for http video streaming,” in Proc. of the
FhMN, 2013, pp. 9–14.
[6] Te-Yuan Huang, Nikhil Handigol, Brandon Heller, Nick
McKeown, and Ramesh Johari, “Confused, timid, and
unstable: Picking a video streaming rate is hard,” in
Proc of the ACM IMC, 2012, pp. 225–238.
[7] Yusuf Sani, Andreas Mauthe, and Christopher Edwards,
“Modelling video rate evolution in adaptive bitrate se-
lection,” in The IEEE International Symposium on Mul-
timedia (ISM 2015), Dec 2015.
[8] John Pastor, Mathematical ecology of populations and
ecosystems, John Wiley & Sons, 2011.
[9] Stefan Lederer, Christopher Muller, and Christian Tim-
merer, “Dynamic adaptive streaming over http dataset,”
in Proc. of the MMsys, 2012, pp. 89–94.
[10] Chao Chen, Lark Kwon Choi, G. de Veciana, C. Cara-
manis, R.W. Heath, and AC. Bovik, “A dynamic sys-
tem model of time-varying subjective quality of video
streams over http,” in Acoustics, Speech and Signal Pro-
cessing (ICASSP), 2013 IEEE International Conference
on, May 2013, pp. 3602–3606.
[11] Lark Kwon Choi, Yiting Liao, and Alan C Bovik,
“Video qoe models for the compute continuum,” E-
LETTER, 2013.
[12] Florin Dobrian, Vyas Sekar, Asad Awan, Ion Stoica,
Dilip Joseph, Aditya Ganjam, Jibin Zhan, and Hui
Zhang, “Understanding the impact of video quality on
user engagement,” ACM SIGCOMM Computer Com-
munication Review, vol. 41, no. 4, pp. 362–373, 2011.
[13] Christopher Mueller, Stefan Lederer, Reinhard Grandl,
and Christian Timmerer, “Oscillation compensating dy-
namic adaptive streaming over http,” in Multimedia and
Expo (ICME), 2015 IEEE International Conference on.
IEEE, 2015, pp. 1–6.
[14] Guibin Tian and Yong Liu, “Towards agile and smooth
video adaptation in dynamic http streaming,” in Proc. of
the CoNEXT, 2012, pp. 109–120.