Adaptive Packet Video Streaming Over P2P Networks Using Active
Measurements
Mubashar Mushtaq and Toufik Ahmed
LaBRI – University of Bordeaux 1
351, cours de la Libération
33405 Talence Cedex
FRANCE
{mushtaq, tad}@labri.fr
Abstract
In this paper we consider the problem of real-time
streaming of IP packet video over Peer-to-Peer
networks (P2P) from multiple senders to a single
receiver. P2P networks are characterized by a
potentially large and highly dynamic population of
hosts that join and leave the network frequently. We
present the design and evaluation of a quality
adaptation streaming mechanism in a multi-source
streaming to a single receiver. Multimedia streaming
is a real time application so, the main challenges in
the design of this mechanism are (1) selection of
senders peers nodes; (2) stream switching among the
peers; (3) optimizing video quality by active
measurements of links; and (4) enhancing the overall
Quality of Service (QoS). Our key technique to
provide quality adaptation is based on active
measurements of network links and selection of sender
peers to enhance the overall throughput. We used
video traffic organized as MDC (Multiple Description
Coding) layers, which provides high error resilient.
Our simulations results using ns2 show that our
solution allows to efficiently utilize available network
bandwidth of sending peers and allow maximizing
streaming qualities at the reception peer.
Keywords: Peer to Peer streaming, Video Streaming,
QoS, Quality Adaptation, Active Measurement.
1. Introduction
Content sharing between communities has
revolutionized the Internet. During the last few years,
we lived a new phenomenon that changed the Internet
business model especially for ISP (Internet Service
Provider). Peer-to-Peer (P2P) systems have gained
tremendous intentions during these years. The Peer-to-
Peer (P2P) phenomenon is facilitating information
flow from and back to the end users. Unlike traditional
distributed systems based on pure client/server model,
P2P networks are self organizing networks that
aggregate large amount of heterogeneous computers
called nodes or peers (nodes and peers are used
interchangeably in this paper). In P2P systems, peers
can communicate directly with each other for the
sharing and exchanging of data, besides this data
exchange these peer nodes also share their
communication and storage resources. The
characteristics of P2P systems make them a better
choice for multimedia content sharing/streaming over
IP networks. P2P systems are dynamic in nature where
nodes can join and leave the network frequently and
that might not have a permanent IP addresses and
observe dynamic changes over the inter connection
links. Virtual networks are built on the top of these
networks at the application level in which individual
peers communicate with each other and share both
communication and storage resources, ideally directly
without using a dedicated server.
The main concept of P2P networking is that each
peer is a client and a server at the same time. P2P
media sharing uses two basic concepts. In the ‘open-
after-downloading’ mode, the media content is played
after downloading all the contents of the file from
different participants, while the ‘play-while-
downloading’ mode allows playing while downloading
the content, which is commonly known as streaming.
The ‘play-while-downloading’ has many advantages
over ‘open-after-downloading’ as it requires lesser
memory and client is not expected to wait for longer
time to finish download first. In this paper, we
consider the problem of Peer-to-Peer streaming
defined as a content streaming from multiple senders
Proceedings of the 11th IEEE Symposium on Computers and Communications (ISCC'06)
0-7695-2588-1/06 $20.00 © 2006
IEEE
to a single receiver in the P2P network, i.e. a single
receiver peer is receiving same content from different
peers present in the P2P network. Multiple sender
peers are selected on the fact that a single sending peer
may not be able or willing to share an outbound
bandwidth of actual playback rate. Dynamic behavior
of P2P systems is another reason of selecting multiple
sender peers for media sharing, as it is possible that
any sender peer sharing media can leave/crash without
any prior notification.
Internet
S1
S2
S4
S3
R
S5
Si
No de sen din g media st ream
(Act ive Node)
Si
Candidate node for sending
media stream
(Can didat e Node)
Node having the content but not
candidate for sending it
R
Node receiving media stream
Figure 1: Peer-to-Peer Multimedia
Streaming Architecture
Figure 1 illustrates our target architecture, which is
composed of many senders and one receiver peer. In
this architecture, several peers are connected logically
and they form P2P network. In figure 1, we present
only those peers which have the requested content. It is
possible that there are many other peers present in the
network which are not having the requested contents
or they don’t intended to share their contents. All the
peers having the requested contents are named
potential peer for sharing of contents. A subset of them
is candidate for sending the content during next period
of time. Furthermore, a subset of these candidate peers
is selected called active peers. The receiver peer
orchestrates the overall streaming mechanism by
selecting potential candidate and active peers. It is
worth noting that the overall quality at the receiver
peer may not increase when additional sender peers are
added because multiple sender peers may be connected
behind the same bottleneck link. For this reason, we
track each peer individually to measure its
performances and capabilities, and then decide
whether to activate sender peer or not, this selection
mechanism is presented in section 4.1.
Real-time traffics are generally carried over Internet
using Real-Time Transport protocol (RTP). P2P
networks are widely used for multimedia streaming.
Quality of the multimedia steams can be affected badly
due to dynamic characteristics of P2P networks.
Quality of video packets is influenced by available
bandwidth, jitter/inter packets delay and packet loss
rate. Inter packet delay/jitter plays major role in
streaming applications. If jitter rate is high there will
be distortion in the video which makes the user
annoyed.
The topology for the construction of virtual P2P
networks and signaling protocol are not considered in
this paper. This topology can be ranged from single
tree approach such as spreadIt, peercast, d3amcat, to
multi-tree approach such as coopnet, splitStream, and
finally mesh-based approach such as Narada and Yoid.
Each approach has its advantages and weaknesses and
we believe that the chosen topology is orthogonal to
our adaptation mechanism that is presented in Section
4. However, peer selection approach is better applied
to a centralized networks embedded in decentralized
networks. This hybrid topology is realized with
hundreds of thousands of peers in the Internet file-
sharing system used in KaZaA [6] and Morpheus [8].
Mostly peers have a centralized relationship to a super
peer called “supernode”. All the queries are forwarded
to this peer (super node) but instead of super nodes
being standalone servers, they band themselves
together in a Gnutella like decentralized network.
Internet, email and SIP proxy also show this kind of
hybrid topology. Mail clients have a centralized
relationship with a specific mail server, but mail
servers themselves share email in a decentralized
fashion. By this way, each super node will have a
matrix containing all peers actually connected to it.
Each peer in the matrix is described by a set of QoS
parameters (for instance available bandwidth, RTT
delay, etc.).
The rest of this paper is organized as follow.
Section 2 presents some related works, Multiple
Description Coding scheme is described briefly in
Section 3. In section 4, we present our proposed
adaptation and Section 5 presents some performance
evaluation and finally, we conclude in Section 6.
2. Related Works
P2P architecture is attracting many researchers and
a lot of research activities are going on, in the domain
of streaming over P2P networks. M. Hefeeda et al. [1]
proposed a mechanism for P2P media streaming using
CollectCast. They proposed an idea for collaborating
with multiple sender peers for media streaming. A
comparison is done for different selection techniques,
i.e. “topology-aware selection” and “end-to-end
Proceedings of the 11th IEEE Symposium on Computers and Communications (ISCC'06)
0-7695-2588-1/06 $20.00 © 2006
IEEE
selection”. In Topology-Aware selection technique all
the shared communication links are considered for best
selection of sending peers while in end-to-end
selection technique these shared segments are not
considered while selection of sender peers. Topology-
Aware selection provides better results because it is
based on congested links monitoring but it offers an
overhead of considering each shared path in the
network.
Reza et al, have proposed a framework PALS [2].
PALS is receiver centric framework, where a receiver
coordinates delivery of layer encoded stream from
multiple senders. A peer selection criterion has been
proposed based on the overall effective throughput.
There is no information available in the start so; initial
peers are selected on random basis. They used layered
coding video for the video transmission.
In [3] Padmanabhan et al. proposed system for the
live and on-demand media streaming using MDC
layers which presents better performance in flash
crowd. Many other researchers have discussed
different aspects of P2P media streaming.
In [4]
congestion control mechanisms using bandwidth
estimation models have been proposed. In [5] a
TCP-Friendly rate allocation algorithm for multiple
sub stream coding combined with path diversity has
been proposed where each sender sends different
streams following different paths.
In this study, we proposed the mechanism of peer
selection which is based on active measurements of
network links. We follow the end-to-end selection
technique. We introduce the cluster approach to avoid
the risks arising from the sharing of same bottleneck
link. The video is composed of MDC layers. We
performed intensive simulations to test the adaptation
mechanism. Results show that our proposed
mechanism improves the overall throughput and
enhance the overall received quality compared to
system without adaptation mechanism. This enhanced
throughput coupled with Multiple Description Coding
improves the QoS remarkably.
3. Multiple Description Coding
Multiple description coding (MDC) [8] and
Layered Coding (LC) are used for Audio/Video
coding. In both schemes each description/layer
contributes one or more characteristics of multimedia
data [10]. Multiple description coding is a method of
encoding the audio and video signals into many
different streams. In Layered coding different layers
are created. Base Layer is one of the most important
layers while all other layers “enhanced layers” are
referenced to base layer. The enhanced layers are not
decodable independently to base layer. In MDC, a sub-
set of descriptions is sufficient to decode the original
file but there will be distortion if number of
descriptions used is very small. By acquiring more
descriptions, distortion can be lowered and it enhances
the overall quality. MDC greatly improves error
resilience because each description can be decoded
independent to other descriptions. Even one
description is sufficient to decode the Audio/Video
signal with its minimum base quality. This feature of
MDC makes it highly applicable for MPEG-4 video
packets transmission over noisy networks/flash
crowded networks when there is more possibility to
loose more video packets. In CoopNet [3] as great
efficiency for the use of MDC has been shown when
used in flash crowd.
4. Adaptive mechanism for P2P packet
video streaming
We are dealing with the problem of unicast, where
a single receiver intended to receive media contents
from many sender peers in P2P network. In this
problem, selection of active peers become more
important as, it is not feasible to select and coordinate
with a larger number of active peers. It leads to extra
overhead of establishing and monitoring of too many
peers and also for reconstruction of all the video
packets before decoding at the receiving end. In the
start of the adaptive mechanism of P2P packet video
streaming, the receiver node sends a query and get
response from sender peers who intend to share the
contents. It’s not necessary that all the nodes having
requested contents must cooperate for content sharing.
It’s a general observation that a large number of nodes
present in P2P networks never intend to share their
resources.
We assume that a lot of peers respond against the
receiver peer query we named these peers as
“Candidate Peers”. Receiver peer selects a sub-set of
candidate peers to start streaming video packets. These
selected peers are called active peers also shown is
Figure 1. The mechanism for peer selection, peer
activation, and stream switching is presented in the
following sub sections.
4.1 Peer Selection Mechanism
Peer selection is an important part of media
streaming in P2P networks as the dynamics and
diversity between peers can vary with the passage of
time which is effected by the facts 1) a sending peer
Proceedings of the 11th IEEE Symposium on Computers and Communications (ISCC'06)
0-7695-2588-1/06 $20.00 © 2006
IEEE
crash/stop contributing the media content: 2) shared
bandwidth is changed: 3) some new peer enter in the
system providing better bandwidth share and low RTT
(round trip time) value: 4) heavy traffic can cause more
packet loss, high inter packets delay which ultimately
causes low QoS (quality of Service). In the result of
these factors, QoS is totally dependent on the
intelligent peer selection and active monitoring of the
network links between peers for detection of said
changing and efficient stream switching to prevent its
effects, stream switching is discussed in section 4.2.
Receiver peer sends a query to search for the desire
media content. In the response, receiver peer maintains
a list of all the candidate peers with whom it can start
streaming. For the selection of a sub-set of candidate
peers, receiver peer diffuses a “Hello” packet to all the
candidate peers. This “Hello” packet serves two
purposes. First it behaves like a ping test to calculate
the Round Trip Time “RTT” between receiver peer
and targeted candidate peers and secondly it gets the
information of targeted peer’s super node. As we
stated, we are considering the P2P architecture where
some peer nodes are connected to one super node to
form a cluster. All the requests pass through these
super nodes. It’s not necessary that this super node is
not acting as server but it is used as transport node.
This super node can be router or switch.
Receiver peer categorize all the candidate peers
according to their “RTT” value and super node index.
Receiver node selects a subset of these candidate peers
having low “RTT” and which are belonging to
different clusters. Selection of candidate peers
belonging to different clusters is justified for the
reason that all the peers present in this cluster (attached
to same super node) share a common bottleneck link,
so it is preferable to choose each sending peer from
different cluster to avoid congestion over same link.
This is an important feature in our adaptation
mechanism.
Different weights w1 and w2 can be assigned for
“RTT” and “TTL” between sender peers. If Pi
represents candidate peers then, sender peer selection
can be made using Eq 1.
RTT * w1
TTL * w2
5
i
|
(Eq. 1)
w1
w2
For this study, we propose the selection criteria
based on “RTT” after performing exhaustive tests to
calculate some performance metrics such as “RTT”
and “number of hops”, details are not presented in this
paper. “RTT” value is used in many mechanisms such
as TCP-Friendly mechanisms and equation-based TCP
throughput.
4.2 Stream Switching Mechanism
P2P networks are not reliable due to their dynamic
nature i.e., any peer can enter or leave the network
without prior notification. There may be problems of
variations in available bandwidth and/or crashing of
peers in the presence of flash crowd. Considering these
problems receiver peer has to monitor all the active
peers regularly for better performance of QoS.
Receiving peer monitor “RTT” value regularly
between all active peers and itself. It’s not desirable to
switch for other candidate peer each time when there is
low “RTT” to avoid oscillating effects. As multimedia
streaming is real time application so for better QoS it’s
always encouraged to have low jitter rate, i.e., inter
packet arriving should be constant and low. Our
adaptation mechanism uses a low-pass filter to
calculate a smoothed value of “RTT”. Bursty traffic
can cause a transient congestion. The “RTT” usage is
not affected by this transient congestion since we
shape this value. The low-pass filter is an exponential
weighted moving average. The EWMA (Exponentially
Weighted Moving Average) Chart is used when it is
desirable to detect out-of-control situations very
quickly. It is an Exponential Smoothing technique that
employs one exponential smoothing parameter to give
more weight to recent observations and less weight to
older observations and vice-versa as presented in the
Eq. 2. When choosing “Ȝ”, it is recommended to use
small values (such as 0.2) to detect small shifts and
larger values (between 0.2 and 0.4) for larger shifts
[11].
X m (1-Ȝ) * RTT + Ȝ * X
(Eq.2)
A buffer is attached at receiver peer and before
playing the media file, a reasonable amount of packets
is received. A threshold value is set for the buffer. The
value depends on the actual playing rate and packets
arriving time. As we proposed MDC for data encoding
,so Receiver node receives different descriptions from
active peers which are decoded after combining to
achieve better quality.
We propose the stream switching for two cases, 1)
if threshold value becomes lesser than that of desired
value (50% of threshold value) receiver node must
look for some other candidate peers. Stream switching
is done by on/off mechanism new candidate peer is
activated (on) sending a request and any of active peer
which has now longer “RTT” can be deactivated (off).
2) For the second case stream switching can be done
when any new peer node enter in the system having
much lesser “RTT” value than that of existing active
peers.
Proceedings of the 11th IEEE Symposium on Computers and Communications (ISCC'06)
0-7695-2588-1/06 $20.00 © 2006
IEEE
n4
n2
n3
n1
n0
Cluster 1
n10
n7
n8
n6
n9
Cluster 2
n15
n12
n14
n11
n13
Cluster 3
n5
n17
n18
R
Cluster 4
n19
MDC1
MDC2
MDC3
MDC1
MDC2
MDC3
MDC3
MDC2
MDC1
CBR Source 1
CBR Sink 1
CBR Source 2
CBR Sink 2
2Mbps
2Mbps
2Mbps
2Mbps
2Mbps
2
M
b
p
s
2
M
b
p
s
Figure 2: Simulation Topology
5 Performance Evaluation
This section presents the simulated results of the
proposed adaptive packet video streaming mechanism.
We performed intensive simulations to validate the
results of our proposed scheme using NS-2 simulator
[12].
5.1 Network Models
The network model considered for simulations is
given in Figure 2. We distributed the original file
equally among different cluster (cluster 1, cluster 2 and
cluster 3). We attached a node “R” in cluster 4 to
received real-time packet video. In this simulation, “R”
tracks “RTT” value between each cluster super node
and itself i.e. “RTT” from n4 to “R”, from n10 to “R”
and from n15 to “R”. This enhances the scalability of
the system rather than tracking each node individually.
Each link in the topology is 2 Mbps bandwidth.
We activate a particular peer in one cluster
depending on “RTT” value. Each sending peer sends
different descriptions of original video file, which are
reconstructed at receiver node “R”. For our test cases,
we have generated 3 different descriptions from
MPEG-4 trace file containing different quality for the
video [13]. These descriptions are generated by the
fractions of DCT (Discrete Cosine Transform) matrix.
We named these descriptions as MDC-1, MDC-2, and
MDC-3. MDC-1 offers 50 % throughput of original
file, MDC-2 offers 40 % throughput of original file,
and MDC-3 offers 30 % throughput of original file.
The overhead caused by MDC coding is about 20% of
the original file. The overall video throughput of the
different MDC layers is given in Figure 3.
We note that no source is providing 100%
throughput but blessing of MDC scheme if receiver
node receives all these three descriptions then it is
possible to reconstruct original file with 100% quality.
We attach all the descriptions to different sender and
add CBR/UDP traffic to overload the network. “CBR
source 1” is sending 512 bytes packet size with 1.5
Mbps. This source is attached to node n2 and sending
UDP datagram to “sink 1” attached to node n7. “CBR
source 2” is same as “CBR source 1” and it is attached
to node n13. It sends UDP datagram to sink attached to
node n19. “CBR source 1” is started at time 5 second,
and stopped at time 55 second. “CBR source 2” is
started at time 10 second, and stopped at time 50. The
duration of the simulation is 60 seconds of time.
To compare the effect of our adaptation mechanism,
we simulate two scenarios, also for making the
scenarios simple, the topology is static. This means
that no peer is leaving or entering the P2P network
during data transfer.
Scenario 1: We run the simulation without applying
any quality adaptation mechanism. In this case, there is
no peer switching done even if a particular peer is
going very congested by CBR traffic.
Scenario 2: We run the simulation with quality
adaptation mechanism by providing peer switching
based on active measurement of “RTT” between the
super node and the receiver node. In this case, “R”
selects and activates sending peers from different
clusters.
5.2 Simulation Analysis
Figure 4 shows the received video traffic at node
“R” in each scenario along with the expected video
quality when using the three MDC layer. As we can
see the adaptation allow maximizing the received
throughput compared to scenario without quality
adaptation. Even with quality adaptation, the received
throughput is less than the expected one since the
heavily stress the network with CBR/UDP traffic. The
CBR traffic causes a lot of packet drops which are
presented in Figure 5. The same comment is applied to
this figure as the packet drop ratio is much lesser in
scenario with quality adaptation compared to scenario
without adaptation. Due to space limitation, we are
unable to present the exact events when the peer
switching is performed.
Proceedings of the 11th IEEE Symposium on Computers and Communications (ISCC'06)
0-7695-2588-1/06 $20.00 © 2006
IEEE