IMPROVING VOD P2P DELIVERY EFFICIENCY OVER
INTERNET USING IDLE PEERS
Leandro Souza, Xiaoyuan Yang Javier Balladini, Ana Ripoll
Computer Architecture and Operating System, Universitat Autònoma de Barcelona, 08193 Bellaterra, Spain
Fernando Cores
Computer Science and Industrial Engineering, Universitat de Lleida, St. Jaume II, 69, 25001 Lleida, Spain
Keywords: On-Demand Media Streaming, Peer-to-Peer systems, Internet VoD.
Abstract: This paper presents DynaPeer Chaining, a peer-to-peer Video-on-Demand (VoD) delivery policy designed
to deal with high bandwidth requirement of multimedia contents and additional constraints imposed by
Internet environment: higher delays and jitter, network congestion, non-symmetrical clients’ bandwidth and
inadequate support for multicast communications. We consider the scenario where we have multiple ADSL-
based peers that stream the same video to multiple receivers. We propose an adaptive scheme to take
advantage of idle peers in order to improve system efficiency, even when extreme conditions (low request
rates or limited peer resources) are considered. We conducted a performance comparison study of our
proposal with classic multicast (Patching) and other P2P delivery schemes, such as P
n
2P
n
and Chaining,
improving their performance by 50%, 62% respectively, even when taking into account Internet constraints.
1 INTRODUCTION
Advances in network technology will provide the
access to new generation, full-interactive and client-
oriented services such as Video-on-Demand.
Through these services, users will be able to view
videos from remote sites at any time. However,
serving multimedia files to a large number of clients
in an “on demand” and “real time” way imposes a
high bandwidth requirement on the underlying
network and server.
To spread the deployment of VoD systems, much
research effort (Guo 2003, Hua 2004, Yang 2005)
has been focused on the delivery process of
multimedia content, exploiting both unicast and
multicast techniques, trying to reduce the bandwidth
consumption and provide better system streaming
capacity. In spite of the success of these techniques,
their scalability requirements to provide service on a
large-scale system, such as Internet, is still limited
by server and network resources.
Recently research has looked to the peer-to-peer
(P2P) paradigm as a solution to decentralize the
delivery process among peers, alleviating the server
load or avoiding any server at all. P2P systems for
streaming video have generated important
contributions. In the Chaining delivery policy (Hua
2004), clients cache the most recently received video
information in the buffer and forward it to the next
clients using unicast channels. The P2Cast (Guo
2004) and cache-and-relay (Jin 2002) allow clients
to forward the video data to more than one client,
creating a delivery tree or ALM. However, neither
Chaining nor ALM delivery policies consider client
output-bandwidth limitation in collaboration
process, which limits their usage to dedicated
network environments. Other VoD P2P-based
architectures such as PROMISE (
Hefeeda 2003),
CoopNet (Padmanabhan 2002) or BitVampire (Liu
2006) assume that a client does not have sufficient
output bandwidth to generate the complete
information to other clients, using n clients to send
the required bandwidth to aggregate. However, they
assume that clients work as proxies storing whole
video information. Furthermore, system scalability is
compromised due to unicast communication. To
solve the scalability problem, in previous works
(Yang 2005) we proposed P
n
2P
n
architecture that
takes advantage of multicast technology on the client
side. This architecture works by exploiting the
297
Souza L., Yang Javier Balladini X., Ripoll A. and Cores F. (2007).
IMPROVING VOD P2P DELIVERY EFFICIENCY OVER INTERNET USING IDLE PEERS.
In Proceedings of the Second International Conference on Signal Processing and Multimedia Applications, pages 293-300
DOI: 10.5220/0002137402930300
Copyright
c
SciTePress
clients non-active resources in two ways: first, it
allows clients to collaborate with the server in the
delivery of initial portions of video, patch streams;
and second, it establishes a group of clients to store
the available information of an existent server
multicast channel to eliminate it. P
n
2P
n
also requires
that output bandwidth is the same as video play-rate.
The Internet environment imposes further
restrictions to P2P streaming schemes in order to
provide VoD service. First, providing service over
non-dedicated network environments implies no
QoS guaranties, transmission congestion, packet loss
and variable point-to-point bandwidth. Second, non-
symmetrical clients’ bandwidth involves a careful
delivery strategy due to clients’ output-bandwidth
limitation. Third, Internet Service-Provider (ISP)
networks differ on supporting (or not) the IP-
Multicast delivery technology. Finally, content
copyright protection affects content storage limited
to non-persistent devices. Thus, content on peers is
only available over a limited period of time.
To solve the above challenges, we have proposed
a VoD architecture for an Internet environment,
which is called P2PVoDSpread. In order to allow
clients (peers) to collaborate with server-delivery
process, we have designed a new delivery scheme
called DynaPeer, based on a P2P paradigm.
DynaPeer policy differs from the previous P2P
schemes in certain key aspects. First, DynaPeer
works with unicast and multicast communication
techniques, depending on the technology available to
the ISP network. Second, this scheme takes into
consideration the non-symmetric characteristics of
client bandwidth, which is in accordance with
current xDSL technology. Third, our delivery
scheme assumes the non-homogeneity features
founded on Internet, which allows us to design a
realistic delivery scheme for VoD services.
In such P2P systems, peer collaboration depends
on clients’ free resources and the availability of
requested multimedia content. If required content is
not available, peers cannot then collaborate, and
their resources will be wasted, becoming idle peers.
In this paper, we focus on a multicast-delivery
policy, DynaPeer Chaining, capable of taking
advantage of idle peers in order to improve
collaboration probability and reduce server load.
This scheme allows the improvement of P2P system
efficiency, even in the most restricted situations
(with low request rates or few peer resources).
Additionally, DynaPeer chaining can operate when
peers’ output-bandwidth is not so restricted,
facilitating further performance improvement.
The remainder of this paper is organized as
follows. In section 2 we present our system
overview. In section 3, the performance evaluation is
presented. In Section 4, we indicate the main
conclusions and future works.
2 SYSTEM OVERVIEW
The goal of P2PVoDSpread design is to take
advantage of client collaboration to decentralize the
server-delivery process, eventually shifting the
streaming load to peers. Clients make their idle-
resources available so as to generate a complete, or
partial, stream for incoming clients.
In this section, we first present details of
P2PVoDSpread system, which is composed by three
entities: VoD Servers, peers and Virtual Servers.
Then we show how system operates, emphasizing on
collaboration window concept, extended buffer
strategy and DynaPeer service schemes.
2.1 System Entities and Concepts
2.1.1 VoD Server
The P2PVoDSpread is not a server-less system;
rather, it combines a server-based architecture with a
P2P delivery scheme. The server holds the entire
system catalogue, acting as seeds for the multimedia
content. It is also responsible for establishing every
client-collaboration process, guarantying the QoS
1
.
2.1.2 Peers
In our system, a peer can be a computer or a set-top-
box, interconnected by ISPs network. It is an active
client who plays a given video and is able to
collaborate with the system.
Peers’ collaboration capacity is limited by peer
resources (bandwidth and storage) and available
video data. In our case, we consider that peers have
an asymmetrical input/output bandwidth (input
bandwidth is, at least, the same as video play-rate
and output bandwidth is supposed to be lower than
video play-rate) and a limited buffer capacity.
Having insufficient output bandwidth to transmit a
complete video stream implies that several peers
have to collaborate in order to provide service for a
complete streaming session. The number of
1
In the rest of paper, the system description is done from point
of view of a VoD system with a single centralized server, which
stores whole system catalogue. However, the system design can
also be directly applied for others architectures composed by
multiple servers (Proxy or CDN based architectures).
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
298
necessary peers to start a stream process to a video i
is denoted as N
i
, and is defined by the ratio between
video i play-rate and peers’ output bandwidth.
Furthermore, due to copyright protection and peers
limited buffer capacity, peers cannot permanently
store a complete video. Therefore, they can only
serve, on the fly, video data previously received
from an active streaming session and temporally
stored on clients’ buffer.
2.1.3 Virtual Server
All collaborations in DynaPeer are managed by the
Virtual Server (VS). A virtual server (Fig. 1),
denoted by VS(i,s,w), is a logical entity defined as a
set of peers that collaborate in a delivery process to
offset s of video i, during a period of time W. The
VS’s service capacity is achieved by peers’ resource
aggregation and will depend on the number of peers
integrating this. The sum of peers’ input (Ij) and
output-bandwidth (Oj) will determine VS input and
output streaming capacity.
Initially, it is assumed that i is the video that all
peers forming VS are reproducing. Video data
available on VS is defined by s (first video block
currently stored on VS) and the collaboration
window W. Outside [s, s+W], the interval defined by
the collaboration window, the VS is unable to make
the collaboration as video data is not available in its
buffer. Therefore, to provide full service for a
streaming session, DynaPeer policies have to
implement a sliding window over whole video data.
In this way, once the collaborative buffer is full, the
following blocks received (s+W, s+W+1, s+W+2,...)
replace the oldest blocks (s, s+1, s+2,...).
2.1.4 Extended Buffer
To take advantage on clients’ buffer capacity,
DynaPeer are configured to set clients’ buffer to
store proportional information of video data that can
be carried out by peers’ output-bandwidth (O
j
). In
this way, peers does not need to store the complete
minutes of video, they are coordinated to store
different blocks of it (i.e., the data kept for future
collaboration for a video i with a play rate Pr
i
, will
be determined by Pr
i
/O
j
relation and buffer capacity
B). We term this strategy as extended buffer. Expr. 1
determines the extended buffer size for a peer j (B
j
).
<
=
otherwiseB
O
O
B
B
ii
i
i
j
,
Pr,
Pr
*
(1)
The extended buffer allows peers to store only
necessary data for collaboration exploiting
efficiently its buffer capacity.
Figure 1: DynaPeer Virtual Server.
2.1.5 Collaboration Window
The collaboration window (W
i
) is defined as the
extended buffer time capacity of all peers involved
in a collaboration process. Due stream patches
requirements the peers need their buffer to store
patching information arising from the ongoing
channel. Thus, extended buffer capacity for
collaboration is more limited, since it can be applied
only in the unused portion of the peer’s buffer
2
. The
W
i,
is calculated by expression 2.
=
Otherwise
BN
B
O
B
NN
B
W
i
ii
i
i
i
i
i
ii
i
i
,
)1*(
Pr
*,
)1(
λ
λ
λ
λ
(2)
The collaboration window defines the maximum
time that an incoming request can be served by peers
(i.e., period of time that any block remains stored on
collaborators’ buffer before its replacement). In this
way, the mandatory condition for the collaboration
process is that the requesting peer arrives inside the
collaboration window.
2.2 System Operationally
DynaPeer operates using Virtual Servers to manage
clients’ collaborations. Each Virtual Server is bound
to an existent ongoing channel and all requesting
peers for this channel, automatically, become
candidate peers inside a new VS in the system.
The VS manages the collaborations by two
different levels: full-stream and partial-stream
collaboration. Full-stream collaboration is achieved
when the VS has sufficient resources to deliver a full
stream to a new client. In this case, the whole video
stream will be delivered by the VS. Otherwise, if
there are not enough resources, the VS proceeds
with the partial stream collaboration. In this case,
VS contributes with the new client request
2
For modeling purposes, we assume the worst collaboration
buffer depending on number of involved peers in stream process
IMPROVING VOD P2P DELIVERY EFFICIENCY OVER INTERNET USING IDLE PEERS
299
proportionally to their service capacity and the
server will be involved in the delivery process,
sending data to the requesting peer in order to
complete the service and to guarantee the QoS. Of
course, every VS begins applying partial-stream
collaboration and when it has sufficient peers and
resources, it switches to full-stream mode.
2.2.1 DynaPeer Multicast
Using the multicast policy, DynaPeer allows the
streaming process for clients in a multi-source/multi-
destination way, better exploiting the network
capacity of ISPs. The VSs are responsible for
creating multicasts channels, from the client’s side
3
,
serving incoming client’s requests. In this way,
DynaPeer avoids any extra server’s resource for
serving contents that have already been started by
other peers.
The collaboration process works by allowing a
new peer joining an ongoing multicast channel
(complete stream) still receiving the entire video
data stream. For new requests for the same video,
the Virtual Server acts in two different ways: First, if
an incoming peer can join an ongoing multicast
channel, the server delivers only the missing portion
of the requested video in a separate unicast channel,
patch stream, using the clients’ output-bandwidth
capacity. The period of time that a peer can join an
ongoing multicast channel is called Patching
Window (detonated as P time), and it depends on
client buffer capacity. Second, if a requesting peer
does not have sufficient buffer space for joining the
ongoing channel (arrival time > P), the VS starts a
new multicast channel for the incoming peer. Once
patching window finishes, DynaPeer begins the
multicast collaborative window (W), whose size
depends on buffer storage available for collaboration
after patching policy. VS only can create a new
multicast channel if the next client request arrives
inside the collaboration window.
Virtual Server will be integrated by all the peers
that arrive inside patching window. Therefore,
depending on a video’s popularity and on clients’
requests rate, it is possible that the number of peers
participating in a VS can be larger than N
i
. In this
case, as only N
i
peers are required to propagate
multicast stream, the remaining VS peers for most
popular videos will not collaborate in the streaming
process. On the other hand, less popular videos VS
cannot have sufficient collaborators peers;
consequently their VS service capacity cannot be
sufficient to fulfill a complete streaming session.
We propose to use the idle peers from over-sized
VS to improve the QoS and performance of VoD
system. In particular, we propose the utilization of
those wasted peers to operate in one of these two
ways: as Backup peers or as Helper peers.
The main goal of a Backup peer is to guarantee
the necessary QoS during a streaming session started
by VS. The Backup peer draws an important rule for
dealing with peer failures and high Internet network
jitter. In this way, Backup peers are used in VS as
support for active collaborators’ peers. If any peers
go down during a streaming session, the Backup
peer will be responsible for replacing it. In addition,
due to bandwidth variations, an active peer inside a
VS, can have its output-bandwidth capacity reduced.
In this case the VS can use the Backup peer to help
on delivery process in order to supply the necessary
quality of service for streaming session.
VS service capacity can be improved by the
utilization of Helper peers. The main function of
Helper peers is to allow the VS of non-popular
videos to achieve full collaboration capacity,
improving DynaPeer performance. Helper peers are
allocated to collaborate with other VS without
sufficient service capacity for carrying out full
stream collaboration. However, Helper peers view
another video and do not have the video data
required to collaborate with different VS. Therefore,
to assist a VS, they previously need to receive video
data, connecting the Helper peer in the ongoing
channel of assisted VS. As result, the Helper peer
downloads video data, proportional to its output-
bandwidth, stores it temporally on collaborative
buffer and uses its output-capacity to delivery it to
another client. The requisite for receiving the data
before serving it will be wasteful unless the ingoing
stream does not require additional resources. This
constraint let this approach feasible only with
multicast communications.
The number of available peers to perform Helper
functionality (H
A
) is achieved after collaboration is
established. This is defined multiplying the total
number of candidate peers inside a collaboration
group (G
i
=B*λ
i
) and that are not involved in the
collaboration process (N
i
) by number of virtual
servers for video i (VS
i
). Expression 3 synthesizes
the number of available helpers in the system, where
M is the total number of videos in system’s catalog.
3
The mechanism of generating multicast trees from clients to
other clients is orthogonal to the analysis presented in this
article. For instance, we assume the mechanism proposed in
(Cheng, K. 2005)
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
300
()()
+=
Di
i
Di
iiiA
CsGCsNGH )1(
>=
jj
NGandMjjD 1|
<=
j
j
jj
WandNGandMjjD
λ
1
1|
(3)
The number of necessary Helpers to participate
in a collaboration process for a video i is defined by
the number of requested peers to complete VS
capacity (N
i
G
i
), multiplied by the number of
complete streams generated for this video, always
provided that the collaboration group needs Helpers
peers (G
i
<N
i
). Expression 4 gives the number of
requested helpers for a video i.
()
<<
=
otherwise
WNGCsGN
H
i
i
iiiii
i
R
,0
1
,)(
λ
(4)
The number of available helpers is limited.
Therefore, we have to decide to which VS the
helpers will be assigned and control when helpers
will be exhausted. To resolve the first issue, we
assign helpers to those VSs that have fewer
requisites to accomplish with a complete stream
capacity, To control the number of available helpers,
we use expr. 3 (available helpers) combined with
expr. 5 that evaluates the total number of helper
peers required by first k videos (more popular):
Fig. 2 shows a snapshot of the system in
multicast configuration. Client arrival rates are
shown in figure (time bar). Peer 1 has sent a video i
request to the server that has started a multicast
channel to attend it. A few minutes later and inside
P
1
time, peers 2, 3, 4 and 5 request the same video.
Theses clients were joined to multicast channel I and
they are incorporated to VS1. In time 2, patching
window finishes and DynaPeer begins the multicast
collaboration window (W
1
). After P
1
time, but also
inside W
1
time, peer 6 requests the same content i
from the server. DynaPeer selects peers 1, 2 and 3
(N
i
=3) to deliver the video and starts a new multicast
channel (channel II) for attending the request. Once
channel propagation is made, peer 4 is set as Backup
peer for VS1 whilst peer 5 is set as a Helper peer.
In time 5, peer 7 requests video i. It arrives inside
P2 time and could be joined to multicast channel II.
In time 8, due time constraints, peer 8 request was
unable to join either multicast channel I or II. The
only possible alternative is to create a new channel.
The VS
1
is unable to create this channel due to its
collaboration window W
1
is surpassed by peer
arrival time.
Regardless that VS
2
was also incomplete in its
total stream capacity to serve the request, it could
achieve its completed service capacity by the
utilization of VS
1
Helper peer 5. At that moment,
VS
2
could start the delivery process to the requesting
peer, generating multicast channel III. Finally peer 9
arrives in minute 9, and it can join the ongoing
multicast channel III
Figure 2: DynaPeer Multicast Snapshot.
2.2.2 DynaPeer Chaining
In DynaPeer, the use of Helpers allows idle clients
to participate in the collaboration process. This
contribution is useful in providing full stream
capacity to a VS, avoiding server participation.
However, depending on request load, not all helpers
can be used by the DynaPeer policy to collaborate
with Virtual Servers. In Fig. 3, we show the number
of total helpers, calculated using expr. 3, and the
remaining free helpers after applying DynaPeer
Multicast (expr. 5). As can be seen, there is an
important volume of free helpers, even with low
request rates. The main idea behind DynaPeer
Chaining is that those free Helpers that are also idle
in the system are grouped to generate a new Virtual
Server in order to propagate video information for a
new period of time. This propagation can be
understood as an extension of a previous VS
collaboration window. Thus, the VS of popular
videos does not need such an extension; in contrast,
however, the VS of unpopular videos does. The
collaboration process works identically to DynaPeer
Multicast. The main difference in this approach is
MjHkH
k
i
i
R
TR
=
=
,)(
1
(5)
IMPROVING VOD P2P DELIVERY EFFICIENCY OVER INTERNET USING IDLE PEERS
301
that there will be special Chain VSs composed only
of Helpers peers.
0
500
1000
1500
2000
2500
1 4 7 1013161922252831343740434649
Request per minute
Helper Peers
Total Helpers
Free Helpers
Figure 3: Free helpers after applying DynaPeer.
Fig. 4 shows a snapshot at minute 12 of
DynaPeer Chaining for an unpopular video. We
have determined that the number of collaborative
clients to attend a full stream capacity is equal to 3.
Furthermore, client buffer are constant and sufficient
to generate two minutes of collaboration window
(W
1
, W
2
and W
3
). There are two requests for this
video, the first starting at minute 0 (peer 1) and the
last at minute 12 (peer 2). When peer 1 makes its
request, a new virtual server is created (VS
1
) and
three Helper peers are designated to it, in order to
guarantee full stream capacity (DynaPeer Helpers).
However, DynaPeer Chaining detects that requested-
video is set as ‘unpopular’, and schedules new
helpers to create a new VS at the final time of the
VS
1
collaboration window (W
1
). This process creates
VS
2
. This new VS will propagate the video for a
longer W
2
period of time. As no requests are
received in this period, DynaPeer Chaining
maintains its function, generating another new VS
(VS
3
) to once again propagate the video data for
future requests. Finally, the request from client 2
arrives at minute 12, being attended by VS
3
.
We notice that, in this example, DynaPeer
Chaining needed 6 Helpers to attend the client’s
request (peer 2). Thus, the number of Helpers for
this purpose will depend on the number of available
Helpers in the system, which is given by the number
of requests for the most popular videos. Thus, with
DynaPeer Chaining, future clients requesting non-
popular videos will have a greater probability of
being attended by a VS, alleviating Server load.
Figure 4: DynaPeer Chaining Snapshot.
3 PERFORMANCE
EVALUATION
In this section, we show the analytical model results
for the DynaPeer delivery scheme (Souza, 2007),
starting by evaluating the performance contrasted
with Patching delivery policy and other P2P-based
delivery policies. The goal of this experimentation is
to analyze DynaPeer Chaining performance
andscalability with respect to different workloads.
The evaluation is based on the server load metric
that is defined as the mean number of streams
required by the server at the end of analysis. We
have evaluated this metric over different workloads
(requests rate) and client’s resources (client-buffer
sizes and output bandwidth). These parameters are
related with each other and affect the peer-
collaboration capacity
3.1 Workload and Metrics
In our experiments, we assumed that inter-arrival
time of client requests follows a Poisson arrival
process with a mean of 1/λ, where λ is the request
rate. We used a Zipf-like distribution to model video
popularity. The probability of the i
th
most popular
video being chosen is
)
1
./(1
1
=
M
j
z
z
j
i
, where M is the
catalogue size and z is the skew factor that adjusts
the probability function. For the whole study, the
skew factor is fixed to 0.729 (typical video-shop
distribution, Hongliang 2006). The time of analysis
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
302
and video length was 90 minutes, the output-
bandwidth of clients is fixed to 750Kbps and video
play rate is 1500Kbps. The analyzed and default
values of the parameters are summarized in table 1.
Table 1: Experimentation environment parameters.
3.2 Client-request Rate Effect
In this experiment, we have changed client request
rate from 2 to 50 requests per minute. The other
system parameters are assumed to have the default
values shown in table 1.
Fig. 5 shows that DynaPeer Chaining performs
DynaPeer Multicast when system load is lesser than
21 req/min. The improvement depends on system
load, but on average it achieves a server load
reduction of 14%. After this load there are not free
helpers for applying DynaPeer Chaining, however,
this fact not impact policy performance, achieving
the same results than DynaPeer Multicast.
DynaPeer policies are capable of accomplishing
the optimal performance of 100 complete streams
(by taking in account initial seed video streams
required for each catalogue video) + patch streams,
with only a request rate of 37 req/min.
Comparing the performance of DynaPeer
policies with other approaches, we can noticed than
as the requests increases, the amount of available
resource of clients also increases, which provides a
lower server-load for DynaPeer policies. DynaPeer
Chaining policy improves Patching requirements by
up to 78%, whilst surpass P
n
2P
n
and Chaining by
50% and 62% respectively.
Figure 5: System Performance vs request rate.
Analyzing the results, we can conclude that
DynaPeer Chaining policy provides an unlimited
scalability, due to it is capable to hold server load in
spite of however the request-rate may have grown.
3.3 Client-buffer Size Effect
The goal of this experimentation is to evaluate the
influence of peer storage resources (buffer size) on
delivery policies performance. To evaluate this, we
have changed client buffer size from 1 to 50
minutes, showing the server-load achieved.
In figure 6, we can perceive that client’s buffer
size have an important impact on system
performance. The server-load of all policies
decreases in accordance with client-buffer size, due
to larger buffer capacity improves sharing
capabilities among requests. The results show that
also DynaPeer Multicast and DynaPeer Chaining
policies are able to achieve better buffer efficiency
than the others. This can be shown by swift
convergence to optimal performance obtained by
DynaPeer policies. DynaPeer achieves lower server
requirements with a buffer capacity of 14 minutes.
Patching achieves its best performance (320
streams) with a buffer capacity for 28 minutes of
video, meanwhile chaining achieves at minute 27.
3.4 Output-Bandwidth Size Effect
In this section, we evaluate the delivery policy’s
performance in accordance with client output-
bandwidth. This parameter only has influence over
DynaPeer policies; due to the other approaches do
not consider client output-bandwidth limitation.
Figure 6: System Performance vs clients’ buffer.
Parameter Default value Analyzed values
Zipf Skew Factor 0.729 0.729
Video length 90 minutes 90 minutes
Video Catalogue Size 100 videos 100 videos
Client’s Output Bandwidt
h
750 Kbps 50 – 2500 Kbps
Request Rate 10 request/min 2-50 request/min
Client’s Buffer Size 5 minutes 1-50 minutes
Play rate 1500 Kbps 1500 Kbps
IMPROVING VOD P2P DELIVERY EFFICIENCY OVER INTERNET USING IDLE PEERS
303
Figure 7: System Performance vs Output-Bandwidth.
Analyzing the figure 7, DynaPeer approaches
have different behaviors. DynaPeer Multicast first
decreases in server-load; when peer output-
bandwidth is over 600 Kbps, it begins to lose
performance and so the server-load increases. This is
caused by the extended buffer definition, which
means higher output-bandwidth creating a lower
extended buffer capacity, and consequently resulting
in lower performance. In this case, better results are
achieved when a tradeoff between extended Buffer
capacity and output bandwidth is employed.
On the other hand, DynaPeer Chaining keeps
alleviating server-load due to helpers’ utilization.
This fact can be possible due the creation of chains
of helpers that can provide a collaboration window
as big as necessary for collaboration. We can infer
that DynaPeer Chaining provides extra capabilities
to adapt P2PVoDSpread system to a heterogeneous
environment. Furthermore, it can take advantage of
additional peer resources (i.e. output bandwidth),
when they are available, to enhance scalability and
performance.
4 CONCLUSIONS
Our concern in this paper is a new VoD system
based on a P2P paradigm and multicast
communication for Internet VoD services. Instead of
independent collaborations between server and
client, the proposed DynaPeer and DynaPeer
Chaining delivery policies synchronize a group of
clients for collaboration to attend new requests,
reducing server-resource requirements.
The experimental study with analytical models
shows that DynaPeer policies improve VoD system
capacity and decrease the server-load, taking major
advantage of client resources to decentralize the
delivery process. Experimental results have shown
that DynaPeer performance depends directly on the
number of available collaborators and their
resources. DynaPeer Chaining can make use of idle
peers to improve system efficiency, even when
extreme conditions (low request rate or limited peer
resources) are considered. Furthermore, it can take
advantage of additional peer resources, to enhance
P2PVoDSpread scalability and performance in an
heterogeneous environment.
We are extending DynaPeer system, analyzing
the impact of dynamic behavior of Internet on our
delivery schemes. We are focused on scheduling and
peer selection policies capable to itself to a
heterogeneous environment. Second, we are working
on a dynamic and distributed control mechanism to
provide fault-tolerance functionality. All these
characteristics will be considered in future work,
using simulation tools and a real prototype.
REFERENCES
Cheng, Kan-L., Cheuk K.-W and Chan S.-H.G., 2005.
Implementation and performance measurement of an
island multicast protocol. IEEE Int. Conf. on
Communications (ICC 2005), vol. 2, pp 1299-1303.
Guo, Y., Suh, K., Kurose J. and Towsley, D.2003. P2cast:
peer-to-peer patching scheme for vod service. In Proc.
of 12th Int. Conf. on World Wide Web, pp 301–309.
Guo, L., Chen, S., Ren, S., Chen, X. and Jiang, S. 2004.
PROP: A Scalable and Reliable P2P Assisted Proxy
Streaming System. In Proc. of the 24th Int. Conf. on
Distributed Computing Systems (ICDCS’04), pp. 778-
786, Washington, DC.
Hefeeda, M., Habib, A., Botev, B., Xu, B., and Bhargava,
D. B. 2003. PROMISE: Peer-to-peer media streaming
using collectcast. In Proc. of ACM Multimedia’ 03,
Berkeley, CA, pages 45-54.
Hongliang Y., Dongdong Z., Ben Y. Z., Weimin Z. 2006.
Understanding user behavior in large-scale video-on-
demand systems In Proc. of the 2006 ACM Eurosys
Conf., Leuven, Belgium.
Hua, K. A., Tantaoui, M., and Tavanapong, W. 2004.
Video delivery technologies for large-scale
deployment of multimedia applications. In Proc. of the
IEEE, volume 92.
Jin, S., and Bestavros, A. 2002. Cache-and-relay
streaming media delivery for asynchronous clients. In
Proceeding of NGC’02, Boston, MA, USA.
Liu, X. and Vuong, S. T. 2006. A Cost-Effective Peer-to-
Peer Architecture for Large-Scale On-Demand Media
Streaming. Journal of Multimedia, Vol. 1, Issue 2.
Padmanabhan, V., Wang, H., Chou, H., and
Sripanidkulchai, K. 2002. Distributing streaming
media content using cooperative networking. In Proc.
NOSSDAV’02, Miami Beach, USA.
Souza, L., 2007. DynaPeer: A Dynamic Peer-to-Peer VoD
System over Internet. PhD thesis, University
Autonoma of Barcelona.
Yang, X. Y., Hernández, P., Cores, F., Ripoll A., Suppi
R. and Luque, E. 2005. Dynamic Distributed
Collaborative Merging Policy to Optimize the
Multicasting Delivery Scheme. In Proc. of 11th Int.
Euro-Par 2005 Conf., Lisbon, Portugal.
SIGMAP 2007 - International Conference on Signal Processing and Multimedia Applications
304