Thisisthepublishedversion:
Luan,TomH.,Kwong,Kin‐Wah,Huang,ZheandTsang,DannyH.K.2008,P2Plivestreaming
towardsbestvideoquality,inIEEECCNC2008:Connectingdevices,c onsumersandcontent:
ProceedingsoftheIEEEConsumerCommunicationsandNetworking2008Conference,IEEE,
Piscataway,N.J.,pp.458‐463.
AvailablefromDeakinResearchOnline:
http://hdl.handle.net/10536/DRO/DU:30063935
Reproducedwiththekindpermissionofthecopyrightowner
Copyright:2008,IEEE
P2P Live Streaming towards Best Video Quality
Hao Luan
∗
, Kin-Wah Kwong
†
,ZheHuang
∗
and Danny H.K. Tsang
∗
∗
Department of Electronic and Computer Engineering
Hong Kong University of Science and Technology
†
Department of Electrical and Systems Engineering
University of Pennsylvania
Email: tomlan@ust.hk, kkw@seas.upenn.edu, ecefelix@ust.hk, eetsang@ust.hk
Abstract—While the overall bandwidth of peer-to-peer live
video streaming system scales automatically as peers collectively
contribute the bandwidth, each peer also demands to download at
the specified video playback rate so as to play the video smoothly.
Therefore, a fundamental problem arisen is how to balance the
bandwidth supply and demand in the peer-to-peer system to enjoy
peers with the best video quality.
To address this problem, we propose a fully distributed peer-to-
peer video streaming framework which automatically adapts the
network towards full bandwidth utilization. Our design possesses
two unique features. First, a special link-level homogenous overlay
network is formed in which all the overlay links approach to have
an identical bandwidth value. With such a feature, video flowing
through the overlay links will not encounter any bottlenecks,
and peers can thus achieve the guaranteed downloading rates.
Second, based on the peer downloading rate observed locally at
the streaming server, the server can adaptively adjust the video
playback rate so that peers can achieve the best video quality with
full bandwidth utilization. The effectiveness of our framework is
verified through extensive simulations.
I. INTRODUCTION
The peer-to-peer video streaming has recently been proposed
as an effective solution on addressing the large scale video
streaming problem. In such a paradigm, peers are allowed
to collaboratively upload the video streams to each other to
contribute the bandwidth. As the pool of overall bandwidth
increases monotonically with more peers subscribing in, a
salient advantage of the peer-to-peer streaming system is that
it scales to support millions of users online at virtually no cost.
However, in the context of bandwidth-consuming multimedia
service, users have the stringent QoS requirements to download
at the specified video playback rate so as to sustain the
continuous video playback. In other words, the live streaming
system must fulfill the bandwidth demand of users given the
limited overall bandwidth collected from distributed users. In
this context, a fundamental problem arisen is how to resolve
the conflict and fully utilize the overall bandwidth to provide
peers with the best delivered video quality.
Bandwidth resource control and allocation are the key com-
ponents in the peer-to-peer video streaming network. However,
due to the heterogenous and dynamic natures of the network,
the problem is quite complicated, and therefore, although it
has indeed been realized, to the best of our knowledge it has
never been seriously studied in previous literature. In face of
the bandwidth constrained video streaming application, Huang
et. al. [5] groups the network into three categories based
A/V Source
Streaming Server
Buffer
Storage
Video Quality
Controller
( )
Scalable Video
Coding (Compress
VideoData with
Playout Rate)
Overlay Network
(Provide peers with guaranteed
downloading rate d )
Peer/Client
Video Decoder
(Playout rate r)
Buffer Storage Media Player
d
Raw Video Data
Channel Coding
(Network Coding)
Channel Decoding
(Network Coding)
r
dr ≈
Fig. 1. Integrated Peer-to-Peer Live Streaming System
on the relationship between the supply and user demand of
bandwidth, say deficit mode, balanced mode and surplus mode
when the overall bandwidth is less than, equal to and more than
the demand, respectively. To construct a feasible network, [5]
argues that the network must work in the balanced or surplus
mode. However, how to control the network to satisfy this
condition has not been discussed in [5]. In [4], Kumar et al.,
show that the network is optimal when the overall bandwidth
supply is fully utilized to fulfill the demand of users in the
balanced mode. In order to achieve this goal, [4] proposes
to perform Call Admission Control (CAC) to select peers
subscribed in according to their bandwidth. However, [4] has
not described any concrete proposals on CAC in peer-to-peer
network.
In this study, we propose an adaptive and fully distributed
system to fully utilize the bandwidth for best-quality video
delivery. As shown in Fig. 1, suppose that the raw video content
is generated at the A/V source and injected into the media
server. Inside the media server, the raw video content is first
compressed by the Scalable Video Coding encoder. The play-
back rate of the compressed video content is determined by the
Video Quality Controller which can sense the downloading rate
of peers and select an appropriate video playback rate accord-
ingly
1
. After compressed, the video content is further encoded
using some channel coding schemes, and then broadcasted to
peers using the peer-to-peer network. At each downloading
peer, the video content is recovered and decoded through the
reverse operations and played at the playback rate. To enable
1
There are many coding schemes, e.g., layered coding with Fine Granularity
Scalability (FGS), to allow us modifying the playback rate and the video quality
in real time.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.
1-4244-1457-1/08/$25.00 © IEEE
458
the adaptive video quality control towards the best delivered
video quality, we encounter the following two challenges due
to the characteristics of the overlay networks.
First, in the large scale distributed system, a huge number of
peers are involved simultaneously. In this case, it is unrealistic
to collect the feedback information from the entire network
and tune the playback rate for individual peers. Therefore, a
key problem arisen is how to determine the appropriate video
playback rate for the entire network in a scalable manner.
Second, to make sure that every single peer can achieve
smooth video playback, a key issue is how to construct the over-
lay network to provide peers with the guaranteed downloading
rate, in a dynamic, heterogenous and distributed network.
In the remaining of this paper, we unfold our journey
by addressing the aforementioned two problems in concrete
details. We summarize our contributions as follows:
• We propose an adaptive framework which can adapt the
playback rate according to available bandwidth of the
network so that all the users are guaranteed with smooth
video playback. While the overall bandwidth is unknown
and dynamically changing all the time with a great number
of heterogenous peers joining and departing the network,
our proposed adaptive system can perfectly balance the
bandwidth demand with respect to the changing and un-
known supply to benefit the overall network.
• We propose a fully distributed system to adaptively plan
the video quality based on the available bandwidth. The
proposed system does not rely on any central information
of the network.
• We construct a special overlay network where all the con-
nections converge to the identical bandwidth. In this case,
the bandwidth resource is equivalent to connections and
can be easily allocated to different peers according to the
QoS requirements. Such a mechanism provides a general
resource allocation method which is also effective in other
peer-to-peer applications, such as BT-like networks [9].
This paper presents its application for live video streaming
which can be considered as max-min resource allocation.
The rest of the paper is organized as follows: Section II
describes the proposed framework in details with analysis. In
Section III we validate our protocol using simulations and
Section IV concludes the paper.
II. P
ROTOCOL DESCRIPTION
We abstract the overlay network as a directed graph G =
{V,E} where V denotes the peer set, including the source node,
and E denotes the link set. In this graph, each node i ∈ V
downloads video streams from multiple parent nodes in parallel
and uploads the streams to multiple child nodes. Let P
i
and
C
i
denote the set of peer i’s parent nodes and child nodes,
respectively. The in-degree of node i is denoted by I
i
=
|
P
i
|
and the out-degree of node i is denoted by O
i
=
|
C
i
|
.LetC
i
denote the uploading capacity of peer i ∈ V.LetN denote the
peer population of the network. Without loss of generality, we
consider one multicast session (or video channel) meaning that
all the peers watch the same video content. The video playback
rate is denoted by r which is a variable and controlled by the
media server as shown in Fig. 1. The media server or source
node is regarded as a normal peer, except that it forever stays
in the network and its in-degree is constantly equal to 0.
Throughout this paper, we make the following two assump-
tions.
First, we assume that Random Linear Network Coding is
used as the channel coding scheme in Fig. 1. In this case,
instead of forwarding the received video streams directly to
downstream nodes, each peer sends the coded data which are
the linear combinations of the incoming video streams [1].
With this scheme, we assume that each node is always able
to retrieve the non-redundant video streams from any selected
parent nodes in the overlay [1].
Second, we assume that the transmission bottleneck is always
on the first hop of the uploading side, rather than inside the net-
work core or on the downloading side. This is due to the widely
adoption of broadband networks and asymmetric access links of
users [2]. In addition, the backbone networks are usually over-
provisioned and optimized using traffic engineering techniques
[10] which make the network core usually congestion-free.
According to the second assumption above, the bandwidth
of fanout connections of a peer i ∈ V can be computed as
C
i
O
i
, with peers evenly allocating bandwidth over the out-going
connections. Using this definition, we propose to construct
a link-level homogeneous overlay network in which all the
overlay connections approach to have identical bandwidth.
In other words,
C
i
O
i
is a constant value same for ∀i ∈ V .A
salient feature of the link-level homogenous overlay is that
video flows along the path of connected connections do not
encounter any bottleneck links, and therefore, peers can achieve
the guaranteed data rates. Moreover, as the bandwidth resource
is now equivalently represented in terms of the number of
overlay connections, we can easily specify the downloading rate
of peers by determining the number of peers’ in-coming con-
nections. To summarize, the link-level homogeneous property
provides peers with the guaranteed end-to-end downloading rate
along the overlay path and enables bandwidth allocation.
A. Overlay Network Construction
In forming a link-level homogenous overlay topology, we
seek to take advantage of the upload bandwidth of all the
participating peers as much as possible and evenly allocate the
transmission workload to peers, i.e., uploading connections.
We construct the network incrementally on the joining and
departing events. Specifically, whenever a peer, e.g., peer j,
requests to join the network or rejoin due to node departures, it
selects an appropriate parent node based on the node’s available
bandwidth and current transmission workloads. As a result, peer
j selects an existing node i ∈ V as the parent node with the
following probability
π
i
=
C
2
i
O
i
∑
x∈V
C
2
x
O
x
=
C
2
i
O
i
Z
(1)
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.
459
where Z =
∑
x∈V
C
2
x
O
x
. This peer selection criterion is inspired by
[7], [8].
By selecting the parent node using Eqn (1), a node with
larger capacity and smaller out-degree in the global network
will be selected with a higher probability. Once this node
is connected, its out-degree will increase by one, making it
unlikely to be selected again by other peers. As new nodes
unceasingly join, the transmission workloads or out-degree of
a peer will finally converge to an equilibrium value which
depends on the peer’s uploading bandwidth and the available
bandwidth of global network. We validate this in the later part
using simulations.
However, selecting peers directly using the probability de-
fined in Eqn (1) requires the global information Z which is not
available. To enable the above peer selection in a completely
distributed manner, we make use of the Markov chain Monte
Carlo (MCMC) using the random walk algorithm.
The idea is to construct a discrete-time ergodic Markov
chain M embedded in the overlay topology G. The states
of the chain M are the nodes in the overlay graph, and the
steady-state distribution of the formed chain M equals to the
desired target distribution defined in Eqn (1). The peer selection
process is accordingly realized using the distributed random
walk algorithm. Specifically, to select a node based on Eqn
(1), a peer first issues a random walker starting at a randomly
selected node (or state) of chain M. The walker is then routed
among connected nodes based on the transition probability
matrix of the Markov chain. The random walk is actually a
process starting from a random initial state of the Markov chain
to converge to the stable state after enough state transitions.
After the random walk process converges, the walker will
stay at a node according to the steady-state probability of the
Markov chain and therefore finishes the selection.
Using the random walk algorithm, we maintain two planes
separately in the overlay network, namely the Markovian plane
and the overlay plane, where the Markovian plane maintains the
Markov chain and is used to route the random walkers only,
while the overlay plane G is the resulting overlay topology
used for video delivery and connected towards the link-level
homogeneity. In order to make the random walk algorithm
converge, the Markov chain imbedded in the Markovian plane
must be ergodic, i.e., aperiodic and irreducible. Therefore, we
construct the Markovian plane to be an undirected graph with
a self-loop connection at each node so that peers are possible
to keep walkers without forwarding to others. In this case,
the formed Markov chain is guaranteed to be aperiodic and
irreducible.
Since the length of the random walks is related to the cost
and accuracy of peer selection, the Markovian plane should
be constructed to minimize the cost of random walks with
the guaranteed accuracy. To achieve this goal, we connect the
Markovian plane as a random regular graph where the degree
of each node is randomly distributed among a fixed region
[n
min
,n
max
], where n
min
and n
max
are predefined constants and
n
min
< n
max
. The efficiency of such a graph is studied in [6].
After forming the graph of the Markovian plane, we still need
to associate the connections with proper transition probability
P =[P
ij
] so that the formed chain converges to the desired
distribution in Eqn (1). To achieve this goal, we invoke the
celebrated Metropolis-Hastings (MH) algorithm. Denote by A
i
the set of peer i’s neighbor nodes in the Markovian plane.
Peer i’ degree in the Markovian plane is therefore |A
i
| + 1,
including the self-loop connection. Given the desired steady-
state probability π = {π
1
,π
2
,...,π
N
} in Eqn (1), P is given as
P
ij
=
1
|A
i
|+1
min
C
2
j
·O
i
·(|A
i
|+1)
C
2
i
·O
j
·
(
|A
j
|+1
)
,1
, j ∈ A
i
, j = i
0, j /∈ A
i
, j = i
1 −
∑
x∈A
i
,x=i
P
ix
, i = j
(2)
where P
ij
is the transition probability from state i to state j.
As shown in Eqn (2), to compute the transition probability, a
peer i only needs to have the local information only, such as
the number of connections in the Markovian plane and overlay
plane respectively and neighbors’ uploading capacity in the
overlay plane.
We are now ready to present the integrated algorithm in
details. Since we maintain two planes in parallel, for each event,
i.e., peer arrival or departure, both G and M should be updated
accordingly.
Joining Procedure: To join the network, a new peer, e.g.,
peer i, first contacts a bootstrap server and fetches a peer list
L
p
. After that, peer i randomly chooses m initial nodes from L
p
and forwards a random walker to each of them. Each walker is
then replayed among peers and traverses over the Markovian
plane based on the local transition probability with Eqn (2)
with the behavior described in Algorithm 1. The walker stops
after TTL steps with each self-loop also counted as one step.
Peers receiving the stopped walkers are selected to upload to
peer i in the overlay topology G.
After updating the overlay graph G, peer i joins the Marko-
vian plane by randomly connecting to
1
2
n
max
peers selected
from L
p
. Then both peer i and the connected peers update their
local transition probabilities using Eqn (2).
Rebuilding Procedure: When a peer departs from the net-
work, each of its child peers in the overlay graph G loses
a parent node. To compensate for the degraded downloading
rate, each of its child nodes acts to reselect a new parent node
Algorithm 1 Random Walk Algorithm
1: Issue a walker to peer x
0
∈ L
p
2: Set n ⇐ 0
3: while n ≤ TTL− 1 do
4: Select a peer x ∈ V with probability P
x
n
x
based on Eqn
(2) where x is either x
n
or a peer x
n
’s neighboring node
in the Markovian plane M
5: x
n+1
⇐ x
6: end while
7: OUTPUT Peer x
n
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.
460
by issuing a walker to rebuild a new link. This behavior is
similar to the joining procedure. In this manner, each node
in the overlay topology is responsible to maintain its own in-
degree to be m, whereas its out-degree is adapted automatically
by the random walk algorithm. As a result, the downloading
performance of each peer is guaranteed by the constant number
of downloading connections it has and the guaranteed data rate
each in-coming connection has.
Meanwhile, when a node departs, all its neighboring nodes
lose a link in the Markovian graph. To repair the graph, if the
nodes’ degree is less than n
min
, they rebuild a new link by ran-
domly selecting a peer in the Markovian plane. Otherwise, they
do nothing. In this process, if a peer in the Markovian plane
has more than n
max
connections, it will deny being connected.
In this manner, the degree of peers in the Markovian plane is
controlled within [n
min
,n
max
]. The pseudocode associated with
above operations is described in Algorithm 2.
Algorithm 2 Formation of Markovian Plane at Peer i ∈ V
INPUT k
i
: Degree of peer i in the Markovian plane
1. Joining Procedure
Send requests and connect to peers randomly
selected in L
p
until k
i
=
1
2
n
max
2. Rebuilding Procedure (When neighboring nodes depart)
if k
i
< n
min
Send connection requests to peers in L
p
and
rebuild new connections until k
i
= n
min
3. Topology Maintenance (On receiving connection requests)
if k
i
< n
max
Accept the request
else Reject the request
Theorem 1: Using the random walk algorithm to select peers
based on Eqn (1), the out-going degree O
i
(t) of a random
selected peer i ∈ V at time t will evolve as
O
i
(t)=
2mC
2
i
H
1 − e
−2(t−t
0
)µ
, t > t
0
0, t ≤ t
0
(3)
where t
0
is the joining time of peer i, µ denotes the mean
departure rate of peers, and H in the denominator is constant.
As a result, the link bandwidth
C
i
O
i
(t)
of a random selected
peer i will converge to the constant δ as
δ = lim
t→∞
C
i
O
i
(t)
=
H
2m
, ∀i ∈ V (4)
We lead the readers to [7], [8] for a detailed proof. Based on
this theorem, the capacity per out-degree value of all the peers
will converge to the identical value, say link-level homogeneity.
In this state, the downloading rate of each peer i converges as
d
i
= δ × m, ∀i ∈ V (5)
B. Adaptive Playback Rate Control
Using our link-level homogeneity overlay, we show that in
the formed network, all the peers converge to the identical
bandwidth per out-degree. As the server also behaves as a
normal peer, its capacity per out-degree value also converges
to the global equilibrium as in Eqn (4). Hence, the server can
estimate the converged downloading rates of peers with Eqn (5)
by observing its own capacity per out-degree value, and based
on this result to adjust the playback rate
2
as follows
r = δ
s
× m ≈ d
i
, ∀i ∈ V (6)
where δ
s
is the capacity per out-degree value of the server node.
The pseudocode is described in Algorithm 3.
Algorithm 3 Playback Rate Adaptation by Source Node s
1: Compute the capacity per out-degree value δ
s
=
c
s
O
s
2: Tune the current playback rate r to be r = δ
s
m
III. EVA L UAT IO N
In this section, we present the evaluation results for our
proposed P2P streaming system. We developed a session-
level, event-driven simulator coded in C++. In each simulation
run, there are totally 50,000 peers inserted into the network
following the Poisson distribution at the mean arrival rate λ.
For each inserted node, we associate it with two parameters:
(1) Life time, which follows the exponential distribution with
the mean
1
µ
. Once the life time of a node expires, the node
is deleted from the network. (2) Node index, which increases
incrementally as nodes are increasingly inserted. The server
node is labeled as node 0 and is always alive with the infinite
life time.
In all the simulation runs, we fix the mean population of
peers N in the stable state to be 10, 000. With the Little’s law,
µ =
λ
N
. In this case, we can use λ to control the churning rate
of the network. By increasing λ, µ will also increase and the
network becomes more dynamic with more peers arriving and
departing in the unit time.
For each experiment, we conduct 10 simulation runs and
compute the averaged value. The default settings of the simu-
lator are as follows: m = 5, n
min
= 20,n
max
= 40,TTL= 10, λ =
10. In this setting, m determines the in-degree and average out-
degree of peers, and therefore, the granularity of the converged
capacity per out-degree of peers as shown in Eqn (4). The
bandwidth distribution of peers is summarized in Table I.
TAB LE I
C
APACITY DISTRIBUTION OF PEERS IN THE SIMULATION
Upload Capacity 256 512 896 2048 5120
Percentage 25% 30% 20% 15% 10%
2
In this paper, we make all the peers have the same downloading rate
by evenly allocate the number of downloading connections to peers. This is
because that all the users require to download at the same playback rate.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE CCNC 2008 proceedings.
461
