Fast Distribution of Replicated Content to Multi- Homed Clients

Short Paper
ACEEE Int. J. on Information Technology, Vol. 3, No. 2, June 2013

Fast Distribution of Replicated Content to MultiHomed Clients
Mohammad Malli
Arab Open University, Beirut, Lebanon
Email: mmalli@aou.edu.lb
Abstract—Clients can potentially have access to more than
one communication network nowadays due to the availability
of a wide variety of access technologies. On the other hand,
service replication has become a trivial approach in overlay
networks to provide a high availability of data and better QoS.
In this paper, we consider such a multi-homed client seeking
a replicated service in overlay network (e.g., CDN, peer-topeer). Our aim is to improve the content distribution by
proposing a new model for being applied at the applicationlevel and in a fully distributed way. Basically, our model
proposes to determine the best mirror server that could be
reached through each client’s network interface based on
application utility function. Then, it consists of downloading
the requested content from the determined best servers
simultaneously through their associated interfaces. Each best
server should deliver a specific estimated range of bytes (i.e.,
content chunk) to an independent TCP socket opened at the
client side for being finally aggregated at the applicationlevel. Our real experiments show that our model is able to
considerably improve the QoS (e.g., content transfer time)
perceived by the client comparing to the traditional content
distribution techniques.
Index Terms—content distribution, service replication, multihoming.

I. INTRODUCTION
Service replication is a scalable solution for the distribution
of digital content over the Internet. The need for this
replication is caused by the increasing number of Internet
users and by the desire to improve the QoS. Also, it is
important for achieving a high availability of data. Many
overlay networks are proposed and installed to realize this
replication: (i) Content Distributed Networks (CDN), where
client requests are forwarded by request redirectors, and
where the contents are stored in mirror servers geographically
distributed over the Internet. Many companies, like Akamai
[1], provide CDNs to content providers. (ii) Peer-to-peer
networks (e.g., bitTorrent [2], where peers behave as clients
and servers. On the other hand, one can profit from multihomed clients to achieve bandwidth aggregation by striping
data across the multiple network interfaces of the clients.
In this paper, we address the problem of improving the
transfer time perceived by multi-homed clients when
requesting digital content replicated in the mirror servers of
one CDN network (resp. peer-to-peer network) or in multiple
ones (i.e., content multi-homing [3], [4], [5]). In the following
discussion, we consider a server as being either a server
among a set of replicated servers in a CDN or a peer in a peerto-peer network that hold the requested content. The best
112
© 2013 ACEEE
DOI: 01.IJIT.3.2.1

server is the one which is able to provide the requested
service to the client with a better QoS than all other servers.
Also, we mean by client a standard client in the client/server
paradigm, or a peer that requests content in a peer-to-peer
network; these terms are used in the paper interchangeably.
Clearly, the best server varies from one client to another based
on many parameters as the performance on the path
connecting the client to the server through each network
interface.
For enhancing the content distribution, many solutions
[6], [7], [8], [9] rely on particular network infrastructure nodes
(e.g., load-aware Anycast router, route controller, peer
coordinator, etc.) taking into account network-specific
constraints (e.g., traffic engineering constraints) perceived
by the ISP or the overlay network operator to solve the
problem as a global optimization problem. While such
approaches could be of great benefit for traffic engineering
purposes, end-systems solutions is able to provide better
enhancement to the performance perceived by the clients.
Besides, in some scenarios, end-systems solutions are able
to achieve a better traffic engineering outcome than the ISPs
can by themselves as shown in [10]. Moreover, one can avoid
the deployment limitations (e.g., network overhead) of the
existing solutions by solving the problem at the end-user
level in a fully distributed way. On the other hand, although
the concept of multipath-capable end systems is interesting
to be applied at the transport level [10], [11], [12], [13], [14],
there is no protocol that simultaneously uses multiple paths
has ever been standardized let alone widely deployed to
replace the most widely used existing protocol TCP.
Therefore, we propose a new model to be applied at the
application-level and in a fully distributed way for improving
the QoS perceived by multi-homed end-users. It consists in
client downloading a replicated content from a certain set of
best mirror servers simultaneously through his/her different
network connections. Firstly, it proposes to determine the
best server that could be communicated through each network
interface based on application utility function. Then, it
consists of downloading the requested content from the
determined best servers simultaneously but with different
estimated amounts.
This must be achieved by opening a TCP connection
with each best server through its associated network interface
to download a specific estimated range of bytes.
The size of this range depends on the weight assigned to
the best server; function of the performance status on its
path to the client’s associated network interface. Thus, our
model is able to improve the QoS perceived by the client

Short Paper
through achieving bandwidth aggregation by striping data
across multiple TCP sockets (i.e., one per network interface)
that download the content chunks from their associated best
servers simultaneously.
This paper is organized as follows. The next section
elaborates the problem of best server selection in replicated
service environment. Then, we present, in Section III, a new
model for distributing replicated content to multi-homed
client. The experiments that show the performance
enhancement provided by this model are presented in Section
IV. Finally, the conclusion is presented in Section V.

perceived in the literature [23]. Hence, these metrics are not
enough to characterize the proximity given the heterogeneity of the Internet in terms of path characteristics and access
link speed, and the diversity of application requirements.
We have realized, in [24], that the proximity must be
characterized in a CHESS space where it is determined at the
application level taking into consideration the network metrics
that decide on the application performance. Therefore, we
have proposed to do that using a utility function that models
the quality perceived by peers at the application level. In this
framework, a peer is closer than another one to some third
peer if it provides a better utility function, whatever the
position of each peer in the geographical and delay spaces.
For example, take the case where the service consists of
clients downloading digital content from a set of replicated
servers using the TCP protocol and where the QoS provided
to clients is maximized if the transfer time is minimized. In this
case, choosing the best server amounts to downloading the
file from the server that is able to provide the minimum transfer
time. This improves the QoS provided to clients and avoids
network and server congestion by distributing the load over
servers and network paths that are less loaded than others.
While the characterization of the proximity in CHESS [24]
has a good impact on application performance, it is a
challenging task due to the two following major requirements.
First, it requires the identification of the appropriate utility
function for each application in a first stage. To solve this
problem, many interesting models have been proposed in
the literature (e.g., transfer time prediction [25], speech quality
prediction [26], [27]). The second challenging task is the
measurement of the different network parameters that impact
the utility function. This is difficult to achieve in large scale
networks where the number of peers can be huge. In such
case, the cost of the direct probing among peers may
outweigh the profit of the characterized proximity. Hence, the
estimation of the network parameters, impacting the utility
function, must be achieved in an easy and scalable way. In
other terms, this should be achieved with a small measurement
overhead and a limited cooperation among nodes. Particularly,
the determination of the network parameters, on the paths
joining a large number of peers, must be achieved in a way
that avoids the direct probing among them as have been
proposed in the literature [15], [16], [17], [18], [19], [20], [21],
[24], [28].

II. SERVICE REPLICATION
Service replication is a scalable solution for the distribution
of digital content over the Internet. The need for this
replication is caused by the increasing number of Internet
users and by the desire to improve the QoS. Also, it is
important for achieving a high availability of the service.
Many overlay networks (e.g., Content Distributed Networks
(CDN), and peer-to-peer networks) are proposed and installed
to realize this replication. The first stage of our approach
consists of determining the best server to be communicated
through each client’s network interface.
Many policies have been studied in the literature for best
server selection. The mostly used approaches can be
classified to the following three categories:
Using the DNS (Domain Name System) to get the IP
address of the best server. This widely used technique
is simple: the DNS servers distribute the IP addresses of
multiple servers associated to a unique name with a round
robin algorithm. It is clear that this solution is not
designed to improve the QoS since it does not consider
any static or dynamic performance limitations. It only
ensures basic load balancing.
Offering the client a list of servers and let him choose
manually the best server to contact. The client choice in
this case is based on his own criteria, for example the
geographical proximity.
 Choosing the closest server in terms of delay. Inferring
the delay closeness between client and servers can be
done using one of the scalable approaches presented in
the literature [15], [16], [17], [18], [19], [20], [21]. Most of
these solutions are based on the network embedding.
Such approaches are based either on network
coordinates or on distance matrix factorization. Also, the
closeness can be determined by identifying the bin of
the client and each server (see [22]). This can be done
by measuring their RTT (Round-Trip Time) to a set of
landmark points. By knowing the bins of the client and
servers, the DNS server can classify the servers (from
the best one to the worst one) based on the distance
between their bins and the client’s one.
Thus, most of the existing solutions for best server selection are based on simple metrics such as the delay, and the
geographical locations which are uncorrelated with other network characteristics (e.g., available bandwidth, loss rate) as
© 2013 ACEEE
DOI: 01.IJIT.3.2.1

III. ENHANCED MODEL FOR DISTRIBUTING REPLICATED CONTENT
IN MULTI-HOMING ENVIRONMENT
A. Proximity Model
The major contribution that we present in this paper is an
extended model of the previously proposed one CHESS [24]
which has been very briefly described in the previous section.
In the new model, we take advantage of the presence of multihoming environment where multiple network connections held
at the peer side to improve the perceived performance when
downloading content from a set of mirror servers. In this
setting, we propose to construct one CHESS space per
113

Short Paper
The size of such chunk depends on the weight assigned to
the best server; function of the performance status on its
path with the client’s associated network interface. Thus, our
model is able to improve the QoS perceived by the client
through achieving bandwidth aggregation by striping data
across multiple TCP sockets (i.e., one per network interface)
that download the content chunks from their associated best
servers simultaneously.
More formally, suppose that the network contains n peers
p ={p1,p2,...,pn} where each peer could play the role of a client
seeking a content or a server holding the requested content.
Obviously, the content is replicated in multiple peers (resp.
mirror servers). The utility function (e.g., delay, available
bandwidth, predicted download time) on the paths joining
peers pi and pj (i,j={1...n}) on top of network connection c
(c={1...k}) is represented by an n X n matrix U c, where Uijc is
the estimated utility function from pi to pj through the network
connection c. The fact that peers could have different number
of network connections and thus different values of k does
not affect the functionality of our model since we are
presenting a distributed algorithm to be executed at each
peer independently. In case that every peer has one network
connection, the system converges to one CHESS space where
the content must be transferred to the client’s unique network
interface from the best server selected as described in Section
II.
Thus, for every multi-homed peer pi,
 the rest of peers pj are ranked through every network
connection c based on the estimations Uijc. We assume
in this model that the larger the utility function value,
the better the quality of service (e.g., available bandwidth
on the network path connecting peers) and the closer
the peers to each other in this space. Obviously, in case
where the utility function is in contrast significant for
small values (e.g., delay, predicted download time), peers
must be ranked according to the increasing order of Uijc.
  its closest peer in the CHESS space c (i.e., best peer
reachable through network connection c) is the peer Pic
that satisfies MAXj={1..n}Uijc (resp.
MINj={1..n}Uijc in
case the utility function is significant for small values).
The best peer Pi c must be different than the ones
determined through the other network interfaces even if
it is the closest peer to pi in the different CHESS spaces.
Thus, if the closest peer in the CHESS space c is the
same one selected as the best peer through another
network connection, then Pic must be selected as the
next closer peer (based on the previously presented
ranking) that is not yet selected as the best peer through
another network connection.
Hence, each best server can upload only one content
chunk to a peer through one of its network connection.
In this way, we are able to pool the capacity over the
space by relying on a good number of best servers (i.e.,
equal to the number of client’s network connections)
instead of uploading the chunks from a fewer number of
best servers (i.e., smaller than the number of client’s

network connection. Thus, one can consider the new model
as multi-dimensional CHESS where peers are ranked in each
CHESS space based on the proximity characterized in this
space; we recall that the proximity characterization is based
on the application utility function.
Then, a peer could position itself in an overlay network
or decide on the size of the content portions to download
from the best peer of each CHESS space based on the proximity
characterization in such overlay networks; our focus in this
paper is on content distribution and not on overlay
construction. Particularly, downloading content from multiple
best servers (i.e., pooling capacity over the space) through
the different network connections simultaneously (i.e.,
pooling capacity over the time) could be achieved obviously
in shorter time delay than that achieved by the trivial pointto-point content delivery techniques. Besides, this can
provide better availability of content and resiliency of service.
To the best of our knowledge, there is no model that pools
capacity over the space and the time as efficient as the one
proposed in this paper although the concept of resource
pooling has been widely elaborated in the literature. Modern
approaches rely on end systems for managing the network
traffic patterns and enhancing the content distribution. In
this scope, many solutions [6], [7], [8], [9] rely on particular
network infrastructure nodes (e.g., load-aware Anycast router,
route controller, peer coordinator, etc.) taking into account
network-specific constraints (e.g., traffic engineering
constraints) perceived by the ISP or the overlay network
operator to solve the problem as a global optimization
problem.
While such approaches could be of great benefit for traffic
engineering purposes, end-systems solutions is able to
provide better performance enhancement perceived by the
clients. Besides, in some scenarios, end-systems solutions
are able to achieve a better traffic engineering outcome than
the ISPs can by themselves as shown in [10]. Moreover, one
can avoid the deployment limitations (e.g., network overhead)
of the existing solutions by solving the problem at the enduser level in a fully distributed way. On the other hand,
although the concept of multipath-capable end systems is
interesting to be applied at the transport level [10], [11], [12],
[13], [14], there is no protocol that simultaneously uses
multiple paths has ever been standardized let alone widely
deployed to replace the most widely used existing protocol
TCP.
Therefore, we take advantage of these perceptions to propose a new model for improving the distribution of replicated
content to multi-homed peer. Basically, it proposes to determine the best server that could be communicated through
each network interface based on application utility function
(i.e., the closest peer in each CHESS space). Then, it consists
of downloading the requested content from the different best
servers simultaneously but with different estimated amounts.
This must be achieved by opening a TCP connection
with each best server through its associated network interface to download a specific estimated range of bytes that
could be considered as a chunk of the requested content.
© 2013 ACEEE
DOI: 01.IJIT.3.2.1

114

Short Paper
network connections). Besides, the content must be
delivered from the selected best servers in a concurrent
way to pool the capacity over the time as well. Moreover,
for more efficient pooling of the servers’ capacity, the
size of a content chunk to be downloaded from each
best server must be proportional to the performance on
its end-to-end path with the associated client’s network
connection (as will be evaluated in this model).
its best CHESS space is the space C that satisfies max
U c . Then, the closest peer to pi in the best CHESS
c={1..k} i(Pi )
space C is denoted by PiC. Thus, PiC is closer to pi than
the other best peers of the rest CHESS spaces since the
estimated utility function between peer pi and PiC has
the greatest value.
Our algorithm can be useful for improving the overlay
construction and content distribution. For overlay construction, a peer could infer its closest peer in each CHESS space
to better position itself in an overlay network; this issue will
be explored and examined in a future work. As for content
delivery purpose which is our focus in this paper, it consists
of determining the correspondent proportion to download
DPic by peer pi through its network connection c from its
associated best server c. Such proportion depends on the
relative value of the performance on the path connecting pi
to Pic through the network connection c with respect to those
on its path with the other best servers reachable through the
remaining network connections. Therefore, we propose to
evaluate DPic as the fraction between the weight wic assigned
to Pic and the overall weights assigned to all the selected
best servers:

(3)
Thus, we propose that peer pi opens a TCP connection with
the best server selected in each CHESS space Pic (c={1...k})
to download the content chunk having the following
proportion from the whole content:

(4)

If these proportions do not divide the file size into finite
ranges, the residual value is added to the range of bytes
allocated to the closest peer in the best CHESS space PiC.
Finally, the receiving application at the client side does the
re-sequencing using a buffer having the size of the requested
content to achieve reliable in-sequence data deliv.ery.
B. Case study
Take the example of Figure 1 where peer p1 would like to
download a content of 999B replicated in peers p2, p3, and p4.
In this scenario, we assume that there are two network
connections c 1 and c 2 that could be used by peers to
communicate with each other. Thus, there are two
correspondent CHESS spaces. Besides, the utility function
is assumed to be a simple metric which is the available
bandwidth on the end-to-end network path for simplicity
seeking. As shown in the figure, the available bandwidth
values on the paths connecting p1 to the other peers through
network connection c1 are:

p1 to p2 is 50 Mbps. So, U12c1 = 50,

p1 to p3 is 25 Mbps. So, U13c1= 25,

p1 to p4 is 10 Mbps. So, U14c1= 10.
For the second CHESS space which is reachable through
the network connection c2, the available bandwidth values
are:

(1)

where,
(2)





and the weight wic evaluates how good is the predicted quality
between pi and Pic comparing to the best case estimated
between pi and PiC. Therefore, one can evaluate it as the
fraction of the utility function evaluated between pi and the
best peer in the CHESS space c (i.e., to be communicated

p1 to p2 is 25 Mbps. So, U12c2 = 25,
p1 to p3 is 50 Mbps. So, U13c2= 50,
p1 to p4 is 100 Mbps. So, U14c2= 100.

In this case, p2 is the best peer for p1 in CHESS space c1
with U12c1 = 50 and p4 is its best peer in CHESS space c2 with
U14c2 = 100. Thus, the best CHESS space for p1 is c2 (i.e., C=
c2).
Then, peer p1 assigns the following weights to the best
servers determined in the two CHESS spaces c 1 and c 2
respectively:
(5)
and,

Figure 1. An example of multi-homed peer seeking a replicated
c on t en t

(6)

through the network interface c) over the utility function
between pi and its closest peer in the best CHESS space:
© 2013 ACEEE
DOI: 01.IJIT.3.2.1

115

Short Paper
Tm which is the latency measured when applying our new
model. In this case, the content is delivered concurrently
from the two best servers identified through the two
network connections according to the multidimensional
CHESS model presented in Section III.
Tb which is the latency measured when downloading the
content in point-to-point mode from the best server
identified through the DSL connection according to the
CHESS model published in [24] and briefly presented in
Section II.
The two ways of content delivery applied in each trace
have been achieved in a sequential manner. The whole traces
have taken place in different dates and times during the month
of July 2012. We assume in these experiments that the utility
function, for best server selection, is the end-to-end available
bandwidth on the network path connecting the client to the
server. One can rely on the predicted transfer time metric [25]
as a more optimized tool in this scope. However, this choice
of utility function does not affect the observations of our
results since our aim is to compare our new model of content
delivery with the classic way of point-to-point content
delivery despite the way of determining the best servers.
Hence, to measure the available bandwidth, we have
applied the probe rate model [29], [30]. This is achieved by
sending a stream of packets from the client to the server at a
rate greater than 512kbps since the end-to-end available
bandwidth is surely smaller (or equal) than this value which
is the maximum download rate of the client per network
interface. Then, the end-to-end available bandwidth is

Thus, peer p1 should download from peer p2 through the
network interface of CHESS space c1 a content chunk having
the following proportion:
(7)
and subsequently the range of bytes [1,333] from the whole
content to download.
Concurrently, peer p1 should download from peer p4
through the interface of CHESS space c2 a content chunk
having the following proportion:
(8)
and subsequently the range of bytes [334,999] from the
whole content to download.
If the file size is 1000B which could not be divided to a
finite ranges of bytes in this case. Then, the residual extra
byte from the division is allocated to the range allocated to
the closest peer p4 in the best CHESS space c2. Then, the
range of bytes to be downloaded from p4 becomes [334,
1000] and the range of bytes to be downloaded from p2
remains the same.
IV. ENHANCED CONTENT DELIVERY PERCEIVED BY THE
APPLICATION
For evaluating the improvement that can be achieved by
applying the presented approach for content distribution,
we have conducted real experiments having the following
settings. We take 10 multi-homed c lients spread in the

Figure 3. Enhancement variation with respect to the file size

calculated as the rate of the receiving stream’s echoes.
Thus, to compare our new content delivery model with
the point-to-point one, we evaluate the enhancement as a
metric having the following expression:

Figure 2. Enhancement of the content delivery time

networks of four ISPs where each client is plugged
simultaneously to two network connections which are DSL
and WiMax. The DSL and WiMax connections have the same
dedicated bandwidth of 512kbps download speed and
128kbps upload speed.
Each trace consists in multi-homed client downloading a
digital content of particular size from one of 20 selected CDN
networks having worldwide distributed sets of mirrors. This
leads to a total of 200 traces where in each trace we measure
the following two content delivery times:
© 2013 ACEEE
DOI: 01.IJIT.3.2.1

(9)
In Figure 2, we plot the CDF (i.e., Cumulative Distribution
Function) of the metric enhancement estimated from the 200
traces. The figure shows that for around 85% of the traces,
our new model for content distribution improves the delivery
time by a proportion between 0.4 and 0.9. The rest of traces
116

Short Paper
(i.e., around 15%) shows an enhancement between 0.1 and
0.4. The considerable enhancement of content delivery time
can be also observed from the expected value of the enhancement metric (averaged over the whole traces) which is
obtained equal to 0.49. In other terms, our traces show that
with high probability, the delivery time could be decreased to
an amount smaller than roughly half of its value when applying our model for content delivery instead of the classic pointto-point delivery method.
To study the variation of the enhancement with respect
to the file size, we plot it in Figure 3 where the first metric is on
the y-axis and the second one is on x-axis. Every point in the
graph represents the average enhancement of several traces
transferring contents of particular size in megabytes. One
can observe from the figure that while the enhancement
fluctuates around the value 0.5, the graph shows a positive
correlation between the enhancement metric and the file size.
Moreover, other measures show that when the file size
increases considerably over 50MB till the value 250MB, the
enhancement fluctuates between 0.6 and 0.7. This interval
of file sizes has not been plotted in Figure 3 since its traces
contain only one trace every increase of 10MB; thus, we are
unable to plot the average enhancement values over this
interval. However, a positive correlation has been also clearly
perceived through the coefficient which obtained equal to
0.47.
Hence, one can realize that our model for content delivery
is able to decrease more considerably the latency when
transferring larger digital contents by striping the data across
multiple TCP sockets (i.e., one per network interface). This
observation can be due to the fact that when the content size
increases the congestion avoidance phase becomes more
dominant than the slow start phase of the download
connection; we notice that the congestion window size
increases exponentially in the slow start phase and then
linearly in the congestion avoidance phase. In this case, it is
obvious to observe better enhancement when reducing
greater number of rounds in the congestion avoidance phase
and spending more rounds in the slow start phase.

applying our model for content delivery instead of the classic
point-to-point content delivery methods. This is due to the
fact that it combines the best server selection scheme with
the bandwidth aggregation facility in multi-homing
environment. Also, the results show that it is able to provide
better quality of service when distributing larger content.
Besides, it is more flexible to be deployed than the
solutions that depend on network infrastructure nodes and
those proposed at the transport layer. Thus, our approach
does not require the deployment of any special network node
and does not impose any change to the existing reliable
transport protocol TCP. This is ensured by working at the
application-level and in a fully distributed way.
Concerning the challenge of determining the best server,
many solutions have been proposed for estimating the
application utility function in a scalable way. This could be
done by firstly defining a function of the parameters impacting
application performance and then relying on a scalable
approach for inferring these parameters using a limited set of
measurements.
Regarding our future work, we will test the efficiency of
the presented model when used for overlay construction.
Besides, we will investigate how our approach reacts to
congestion.
ACKNOWLEDGMENT
The author wishes to thank Mr. Ahmad Moosa for his
valuable help in the testing phase.
REFERENCES
[1] Akamai, http://www.akamai.com.
[2] BitTorrent, http://www.bittorrent.com.
[3] H. H. Liu, Y. Wang, Y. R. Yang, H. Wang, and C. Tian,
“Optimizing cost and performance for content multihoming”,
ACM SIGCOMM, 2012.
[4] V. K. Adhikari, Y. Guo, F. Hao, M. Varvello, V. Hilt, M. Steiner,
and Z.-L. Zhang, “Unreeling netflix: Understanding and
improving multi-CDN movie delivery”, IEEE INFOCOM,
2012.
[5] G. Bertrand, E. Stephan, G. Watson, T. Burbridge, P. Eardley,
and K. Ma, “Use cases for CDNi”, IETF Draft, 2012.
[6] H. A. Alzoubi, S. Lee, M. Rabinovich, O. Spatscheck, and J.
Van Der Merwe, “A practical architecture for an anycast CDN”,
ACM Transactions on the Web, vol. 5, pp. 17-29, 2011.
[7] R. S. Peterson, B. Wong, and E. G. Sirer, “A content propagation
metric for efficient content distribution”, ACM SIGCOMM,
2011.
[8] I. Poese, B. Frank, B. Ager, G. Smaragdakis, and A. Feldmann,
“Improving content delivery using provider-aided distance
information”, ACM IMC, 2010.
[9] R. S. Peterson and E. G. Sirer. Antfarm, “efficient content
distribution with managed swarms”, NSDI, 2009.
[10] D. Wischik, M. Handley, M. Braun, “The resource pooling
principle”, ACM SIGCOMM CCR, vol. 38, pp. 47-52, 2008.
[11] D. Wischik, M. Handley and C. Raiciu, “Control of multipath
TCP and optimization of multipath routing in the Internet”,
NetCOOP, 2009.
[12] M. Zhang, J. Lai, A. Krishnamurthy, L. Peterson, and R. Wang,
“A Transport Layer Approach for Improving End-to-End

V. CONCLUSIONS AND PERSPECTIVES
In this paper, we propose a new model to improve the content
distribution in overlay networks. Our model takes advantage
from the content replication and multi-homing facilities which
are widely available nowadays. This is done by determining
the best peer that could be reached through each network
interface based on the estimation of the application utility
function. Then, it consists of achieving bandwidth
aggregation by striping data across multiple TCP sockets
(i.e., one per network interface) that download the content
chunks from their associated best servers simultaneously.
Our extensive real measurements show clearly how our
solution outperforms the existing solutions by decreasing
considerably the content distribution time. Our traces show
that with high probability, the delivery time could be decreased
to an amount smaller than roughly half of its value when
© 2013 ACEEE
DOI: 01.IJIT.3.2.1

117

Short Paper

[13]

[14]

[15]
[16]
[17]

[18]

[19]
[20]
[21]

Performance and Robustness Using Redundant Paths”,
USENIX, 2004.
H. Hsieh and R. Sivakumar, “pTCP: An End-to-End Transport
Layer Protocol for Striped Connections”, IEEE ICNP, Paris,
2002.
H. Hsieh and R. Sivakumar, “A Transport Layer Approach
for Achieving Aggregate Bandwidths on Multi-homed Mobile
Hosts”, ACM MOBI-COM, Atlanta, GA USA, 2002.
E. Ng and H. Zhang, “Predicting Internet network distance
with coordinates-based approaches”, IEEE Infocom, 2002.
L. Tang, and M. Crovella, “Virtual Landmarks for the Internet”,
ACM IMC, 2003.
B. Wong, A. Silvkins, and E. G. Sirer, “Meridian: A
Lightweight Network Location Service without Virtual
Coordinates”, ACM SIGCOMM, 2005.
P. Francis, S. Jamin, C. Jin, Y. Jin, D. Raz, Y. Shavitt, and
L.Zhang, “IDMaps: A Global Internet Host Distance
Estimation Service”, IEEE/ACM Transactions on Networking,
2001.
Y. Shavitt, and T. Tankel, “Big-Bang Simulation for Embedding
Network Distances in Euclidean Space”, IEEE Infocom, 2004.
H. Lim, J. Hou, and C. Choi. “Constructing Internet Coordinate
System based on Delay Measurement”, ACM IMC, 2003.
J. Ledlie, P. Gardner and M. Seltzer, “Network Coordinates
in the Wild”, NSDI, 2007.

© 2013 ACEEE
DOI: 01.IJIT.3.2.1

[22] S. Ratnasamy and M. Handly, R. Karp and S. Shenker,
“Topologically-Aware Overlay Construction and Server
Selection”, IEEE Infocom, 2002.
[23] M. Malli, C. Barakat, and W. Dabbous, “Application-level
versus Network-level Proximity”, Asian Internet Engineering
Conference, Thailand, 2005.
[24] M. Malli, C. Barakat, and W. Dabbous, “CHESS: An
Application-aware Space for Enhanced Scalable Services in
Overlay Networks”, IEEE Computer Communication Journal,
vol. 31, pp. 1239-1253, 2008.
[25] M. Malli, C. Barakat, and W. Dabbous, “An Efficient
Approach for Content Delivery in Overlay Networks”, IEEE
CCNC, 2005.
[26] L. Ding, and R. Goubran, “Speech Quality Prediction in VoIP
Using the Extended E-Model”, IEEE Globecom, 2003.
[27] R. Cole, and J. Rosenbluth, “Voice over IP Performance
Monitoring”, ACM SIGCOMM CCR, vol. 31 , pp. 9-24, 2001.
[28] N. Hu, P. Steenkiste, “Exploiting Internet Sharing for Large
Scale Available Bandwidth Estimation”, ACM IMC, 2005.
[29] N. Hu and P. Steenkiste, “Evaluation and Characterization of
Available Bandwidth Techniques”, IEEE JSAC Special Issue
in Internet and WWW Measurement, Mapping, and Modeling,
2003.
[30] V. J. Ribeiro, R. H. Riedi, R. G. Baraniuk, J. Navratil, and L.
Cottrell, “PathChirp: Efficient Available Bandwidth
Estimation for Network Paths”, Passive and Active
Measurement Workshop, 2003.

118

Fast Distribution of Replicated Content to Multi- Homed Clients

More Related Content

What's hot (19)

Viewers also liked (8)

Similar to Fast Distribution of Replicated Content to Multi- Homed Clients (20)

More from IDES Editor (20)

Recently uploaded (20)

Fast Distribution of Replicated Content to Multi- Homed Clients