A Survey of BitTorrent Performance

Abstract

Since its inception, BitTorrent has proved to be the most popular approach for sharing large files using the peer-to-peer paradigm. BitTorrent introduced several innovative mechanisms such as tit-for-tat (TFT) and rarest first to enable efficient distribution of files among the participating peers. Several studies examining the performance of BitTorrent and its mechanisms have been published in the literature. In this paper, we present a survey of performance studies of BitTorrent from 2003 to 2008. We categorize these studies based on the techniques used, the mechanisms studied and the resulting observations about BitTorrent performance. Many of the performance studies also suggested modifications to BitTorrent's mechanisms to further improve its performance. We also present a survey of the suggested improvements and categorize them into different groups.

Full text

140 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
A Survey of BitTorrent Performance
Raymond Lei Xia and Jogesh K. Muppala, Senior Member, IEEE
Abstract—Since its inception, BitTorrent has proved to be
the most popular approach for sharing large files using the
peer-to-peer paradigm. BitTorrent introduced several innovative
mechanisms such as tit-for-tat (TFT) and rarest first to enable
efficient distribution of files among the participating peers.
Several studies examining the performance of BitTorrent and
its mechanisms have been published in the literature. In this
paper, we present a survey of performance studies of BitTorrent
from 2003 to 2008. We categorize these studies based on the
techniques used, the mechanisms studied and the resulting obser-
vations about BitTorrent performance. Many of the performance
studies also suggested modifications to BitTorrent’s mechanisms
to further improve its performance. We also present a survey of
the suggested improvements and categorize them into different
groups.
Index Terms—BitTorrent, Large file downloading, Peer-to-peer
Networks.
I. INTRODUCTION
THE PEER-TO-PEER (P2P) paradigm has proved to be
a very effective approach for designing scalable and
robust networking applications. This approach overcomes the
scalability limitations of the traditional client/server approach.
Large scale file-sharing and file-distribution applications were
the first to exploit the new paradigm. BitTorrent [1], one of
the most popular peer-to-peer (P2P) file distribution applica-
tions, was introduced by Bram Cohen in 2001 [2]. BitTorrent
was quickly embraced by several content providers such as
Lindows, Blizzard and most Linux distributions as a scalable
way of delivering content, decreasing the load on congested
servers, minimizing the distribution cost and reducing the
downloading time for users. BitTorrent is a second generation
P2P file sharing system [3], which focuses on organizing the
peers that are interested in downloading the same file into an
overlay network, called a Torren t . Interestingly, as more users
join a Torrent, the downloading rate that is achieved by all
the peers increases. BitTorrent is noted for introducing several
innovative mechanisms including tit-for-tat (TFT) and rarest
first (RF) that it uses for cooperative file distribution among
the participating peers.
Any file distribution application implemented using the P2P
paradigm requires two different functions to be supported: (1)
a search function that enables peers to locate the content that
they are interested in among the participating peers, and (2)
Manuscript received 17 June 2008; revised 14 December 2008. The work
described in this paper has been supported by HK RGC under RGC-
Competitive Earmarked Research Grant HKUST 617907
R. L. Xia is with the Dept. of Computer Science and Engineering, The Hong
Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong
Kong (e-mail: xialei@cse.ust.hk).
J. K. Muppala is with the Dept. of Computer Science and Engineering,
The Hong Kong University of Science and Technology, Clear Water Bay,
Kowloon, Hong Kong (e-mail: muppala@cse.ust.hk).
Digital Object Identifier 10.1109/SURV.2010.021110.00036
a downloading function that enables the peer to download
the content once it is located. Traditional P2P file sharing
systems such as Gnutella [4] and KaZaa [5] focus on quickly
locating the peers that possess a given file. Once located,
the file is directly downloaded from one of the peers having
the file. BitTorrent, on the other hand relies on cooperation
among the peers for file distribution using swarming [2].
BitTorrent differs from the conventional P2P file sharing
systems in three aspects. First, BitTorrent application does not
provide a search function. Instead, peers are expected to peruse
central-directory based search facilities provided by several
supporting websites. Second, it provides a file-level sharing
policy instead of a directory-level sharing policy. Finally, it
employs a bartering mechanism such as TFT among peers [6],
such that unless a peer contributes to the ongoing downloading
of file pieces, it will not be able to download the file.
The widespread popularity of BitTorrent has attracted the
attention of several researchers who conducted various per-
formance studies in order to understand the behavior of the
BitTorrent protocol, its mechanisms and the overall application
performance. These studies affirmed the importance of TFT
and rarest first mechanisms for BitTorrent’s success. At the
same time, several of these studies identified the impact of
BitTorrent’s phenomenal success on the traffic in computer
networks. For example, one of the significant problems is the
increase in file downloading traffic in the Internet. Reports
show that BitTorrent has become the largest source of P2P
file sharing traffic in the Internet: BitTorrent traffic accounted
for 18% of broadband traffic in 2006 [7]. Karagiannis et al.
[8] also report that BitTorrent occupies about 13-15% of the
access link bandwidth at a residential university. Furthermore,
BitTorrent presents significant traffic-engineering challenges
for Internet service providers (ISPs). Current implementations
of BitTorrent ignore the underlying Internet topology or ISP
link costs and result in a large amount of cross-ISP traffic
[9]. These issues show that BitTorrent protocol has further
scope for improvement in its core mechanisms. For example,
one of the most criticized of BitTorrent’s mechanisms is the
random neighbor selection strategy, which prolongs the file
downloading time and results in inefficient usage of network
resources. Some researchers also suggested BitTorrent-like
protocols [10], [11], [12], which modify the BitTorrent mech-
anisms to further improve performance.
In this paper, we present a comprehensive survey of all
the performance studies on BitTorrent. We summarize the
salient findings of these performance studies. We then present
a survey of all the improvements that have been suggested
for BitTorrent and its mechanisms by various researchers. We
categorize these improvements into various groups so as to
enable us to understand the interplay among the suggested
1553-877X/10/$25.00 c
2010 IEEE
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 141
improvements. Surveys of P2P networks in general, like
[13] and [14], application-layer multicast protocols [15], P2P
security issues [16], and P2P based resource discovery [17]
are already available in the literature. The difference between
these surveys and our paper is that we focus specifically on
BitTorrent, which is one of the most popular P2P applications.
The remainder of this paper is organized as follows. In
section II, we present an overview of BitTorrent and its
mechanisms. In Section III, we present a survey of the studies
on BitTorrent performance. Section IV reviews several of the
improvements suggested for BitTorrent mechanisms to further
improve its performance. Then we compare the suggested im-
provements and categorize them into several groups. Finally,
we conclude this paper in section V.
II. OVERVIEW OF BITTORRENT
A. What is BitTorrent
As stated earlier BitTorrent (BT) is one of the most popular
applications for distributing large size files. The application
is implemented as a hybrid P2P system, with most of the
interaction directly among the peers, but requiring occasional
interaction with a server for locating peers. The BitTorrent
protocol requires peers to organize themselves into an overlay
network, with connections among the peers, for each file being
distributed. This overlay network is called a To rren t . Each file
being distributed by BitTorrent requires the establishment of
separate torrent. Besides the peers, a tracker and a web server
play an important part in file distribution using BitTorrent.
The tracker is a special infrastructure node which stores
meta-information about the peers that are currently active
within a torrent. Peers interact with the tracker using a simple
protocol layered on top of HTTP in which a peer sends
information about the file it is downloading and the port
number [2] to the tracker. The tracker does not participate in
the actual distribution of the file, but only serves the purpose
of enabling peers to find each other.
The peers that are part of a torrent can be classified into
two types: a seed and a leecher. A seed is a client that has a
complete copy of the file and remains in the torrent to serve
other peers. For a torrent to get started, we need at least one
initial seed that provides the entire content for download. A
leecher is a client that is still downloading the file.
When a file is to be made available for distribution through
a torrent, then a special file with a .torrent extension is
made available on a web server. The .torrent file contains
information about the file including its length, name, hashing
information, and the url of a tracker [2]. A peer that wants
to download a file first obtains the corresponding .torrent
file from the web server. Then, the peer contacts the tracker
and requests a list of IP/port pairs of other peers that are
already participating in the torrent. When the tracker receives
a request, it will send the requesting peer a list of potential
neighboring peers, usually 50, that are selected randomly from
the set of all active peers within the torrent. After obtaining the
list, the peer contacts around 20–40 peers from the list to add
them as its neighbors. This set of neighbors forms the peer set
for the joining peer. If the number of neighbors connected to
a peer falls below 20, the peer will again contact the tracker
Peer A
[Leecher]
Peer C
[Seed]
Peer B
[Leecher]
Tracke r
Web Server
123
5
4
Fig. 1. BitTorrent file sharing process
and obtain a new list. The data flow between two connected
peers is bidirectional over the connection.
The steps in file sharing using BitTorrent are shown in
Fig. 1. There are a total of five steps:
1) When peer A wants to join a torrent, it first downloads
the corresponding .torrent file from a web server.
2) Then, peer A contacts the tracker and asks for a list of
peers that are already participating in the torrent.
3) Next, the tracker sends a list of about 50 peers which
are participating in the torrent.
4) Then, peer A selects 35 peers from the list as its
neighbors, and establishes connections with them.
5) After the connections are established, peer A can ex-
change file pieces with the neighbours.
The file exchange among peers uses a swarming technique
[18], in which the file is broken into fixed size pieces (typically
256 KB each), and peers exchange pieces with each another.
When a piece is completely downloaded, its SHA1 hash
is computed and compared with the value in the metafile
(.torrent). If the values match, the peer will announce the
availability of the complete piece to its neighbours.
BitTorrent uses pipelining in order to keep its TCP con-
nections operating at full capacity. To facilitate this, a piece
is divided into sub-pieces, typically of 16 KB in size, which
are called blocks or chunks. Requests for sub-pieces are kept
pending in a pipeline, typically five in number. Each sub-piece
arrival will trigger a request for another sub-piece. This way
the connection is kept busy most of the time.
A peer downloads pieces not only from the seed(s), but
also from other peers, thereby substantially reducing the
load on the seed(s). A peer usually can serve four peers
simultaneously, and it chooses the best four peers to unchoke
and chokes other requesting peers. Choking is a temporary
refusal to upload, but the connections are not closed.
B. Mechanisms of BitTorrent
The mechanisms employed by BitTorrent mainly consist
of peer and piece selection strategies. A good peer selection
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

142 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
A’s Peer Set
Interested Set
Mutually Interested Set
Regular
Unchoke
Set Optimistic
Unchoke
Set
Fig. 2. Peer Sets
strategy should maximize the service capacity of the system,
and an efficient piece selection strategy should guarantee that
each peer can find pieces that are interesting to it from its
neighbors. A detailed description of these mechanisms is
given in [2]. Here we briefly summarize these mechanisms
for completeness so that further discussions in the subsequent
sections can be placed in proper context.
As already mentioned in the previous section, the peer set
of a peer A is the set of all the neighbors of A. Peer A will
consider a peer B belonging to its peer set as an interested
peer, if Peer B has pieces that A does not possess. We can
define an Interested Set of Peer A as the set of interested
peers, which is a subset of its peer set. A peer can continue
its download only if its interested set is non-null. Ideally Peer
A should use appropriate mechanisms to maximize the size
of its interested set. Doing so increases the ability of Peer
A to download file pieces from the maximum number of
other peers. Similarly, Peer A considers peer B as a mutually
interested peer if and only if peer B is in peer A’s interested
set and peer A is also an interested peer for Peer B, i.e., Peer
A has pieces that Peer B does not possess. It is beneficial to
maximize the size of the mutually interested set because this
increases the chances for trading pieces with the neighbors,
thus effectively increasing the downloading rate.
A peer A considers a peer B as a member of its regular
unchoke set if peer B is unchoked by Peer A using the
TFT mechanism. Peer A considers peer B as a member of
its optimistic unchoke set if peer B is unchoked by Peer A
due to the optimistic unchoking mechanism. Details of these
mechanisms are given later in this section. We illustrate these
sets and their relationships further in Fig. 2. These sets play
an important role in determining the downloading experience
of any peer.
1) Peer Selection Strategy: The peer selection strategy uses
four main mechanisms: tit-for-tat (TFT), optimistic unchoking
(OU), anti-snubbing, and upload only. Summed up together,
they form the basis for the choking algorithm used by a
peer. The major aim of these mechanisms is to improve the
downloading experience of those peers that contribute to the
file exchange by uploading pieces to others, and punish free-
riders, i.e., those peers that download pieces from others
without uploading any pieces.
•Tit-for-tat: The tit-for-tat strategy is used by a peer to
preferentially upload to its neighbors who provide it
the best downloading rates. At any given time, a peer
exchanges file pieces with a fixed number of neighboring
peers, usually at most four. A leecher preferentially
uploads to the best three neighbors who provide it the best
downloading rate and chokes others. Every 10 seconds,
the leecher reevaluates the downloading rate from all
peers who are sending data to it. If another peer offers
better downloading rate, the leecher will choke the peer
with the smallest downloading rate among the current
peers it is serving, and unchoke the better peer. This
mechanism is very important to encourage contributors
and punish free-riders, thus preventing leechers from
downloading without contributing anything.
•Optimistic Unchoking: Besides the TFT mechanism,
BitTorrent incorporates an optimistic unchoking policy.
Every 30 seconds, a peer unchokes a neighboring peer
randomly regardless of its uploading rate. The benefitof
Optimistic Unchoking is that it enables a peer to discover
better neighboring peers to exchange file pieces with
since other neighbors may have higher capacity for up-
loading than the currently unchoked peers. If a peer uses
only the TFT mechanism, there will be no opportunity for
discovering other peers that can provide higher uploading
rate. In addition, this strategy is especially useful for
newly joined peers to get started.
•Anti-snubbing: A peer may be choked by the peers it
was formerly downloading from, thereby getting poor
downloading rates. To address this problem, when a peer
notices that some time has elapsed without getting a
single piece from a neighboring peer, the leecher assumes
it is ’snubbed’ by that peer and does not upload to it any
further through regular unchoke.
•Upload Only: Once a peer finishes downloading the entire
file, it becomes a seed. Since seeds have nothing to
download, they cannot select peers based on downloading
rates. Instead, the seeds prefer to upload to peers with
better uploading rates [2]
2) Piece Selection Strategy: The second major strategy that
plays an important role in the success of BitTorrent is the piece
selection strategy. When a peer wants to download pieces from
its neighbors, it adopts several piece selection strategies, which
includes the following four mechanisms:
•Strict Priority: In BitTorrent, peers concentrate on down-
loading a whole piece before requesting another piece.
Thus if a sub-piece is requested, then subsequent sub-
pieces of the same piece will be requested preferentially
in order to complete the download of the whole piece as
soon as possible, because only complete pieces can be
traded with others.
•Rarest First: Peers often prefer to download pieces which
are the rarest among their neighbors [2]. This strategy is
called the rarest first algorithm which works as follows.
Each peer maintains a list of the pieces that each of its
neighbors possess. This list is updated each time a copy
of a piece becomes available from its neighbors. It then
constructs a rarest-pieces set which is the list of those
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 143
pieces that have the least number of copies among its
neighbors. It then selects the piece to download that is
rarest among its neighbors. This strategy increases the
chance of a peer trading with its neighbors as it has pieces
that others want. This also benefits the seeds by reducing
their server burden, since they need to serve out less
copies of the same piece. Furthermore, the departure of
a seed will not leave the remaining downloaders without
the opportunity to exchange file pieces, because this
strategy ensures that all the pieces are quickly distributed
to at least some of the leechers [2].
•Random First Piece: There is an exception to the rarest
first policy when a peer first joins a torrent. Since the peer
does not have any pieces, so it would like to download a
whole piece quickly so that it can get ready to reciprocate
with the TFT algorithm. In this case, the new peer will
not seek the rarest piece. Instead, it will download the
first piece randomly in order to make sure it can get a
whole piece as fast as possible.
•Endgame Mode: This mode is adopted by a peer towards
the end of downloading the file. If a piece is requested
from a peer that has a slow transfer rate, the downloading
time will be prolonged. To address this problem, a peer
requests all of its neighbors for blocks it has not received.
Once a block is obtained, the peer cancels the request for
that block from its neighbors so as to decrease bandwidth
wastage due to redundant downloads.
C. BitTorrent-like Systems
Some researchers proposed BitTorrent-like systems such
as Slurpie [12], Fox [11], and Avalanche [10]. They have
a lot of similarities in their operations with BitTorrent. The
common goal of these approaches is to improve the file sharing
performance.
Slurpie [12] is designed with the goal of reducing the
clients’ downloading time and making the client sets scalable.
In Slurpie, when a node wants to download a file, it first
contacts a server called a topology server. The topology server
sends the node a list of other nodes which are downloading
the same file. The node then contacts the other nodes in the
list to form a mesh. Similarly, the file is divided into fixed
size blocks, and is exchanged through the mesh. The authors
claim that Slurpie has advantages in adapting to varying
bandwidth and achieves better scaling with the number of
neighbors compared to BitTorrent. Slurpie employs an effec-
tive downloading bandwidth estimation technique described
in [12] which reports three different states of bandwidth for
each node: underutilized, at-capacity, and throttle-back. Each
node in Slurpie maintains a target number of neighbors, and
this number is continually updated according to the available
bandwidth. The bandwidth estimation algorithm runs every
second, and if the node’s state is underutilized and the node
needs more blocks, a new connection will be opened. By using
these strategies, Slurpie can improve the system performance
as the size of the network increases.
Levin et al. [11] proposed a Fair, Optimal eXchange proto-
col, called FOX to achieve a fair file swarming. FOX provides
not only an effective application of the TFT mechanism,
but also theoretically optimal downloading time for peers. In
FOX, the authors assumed that all peers are greedy. In the
overlay structure of FOX, there are Nclients downloading
afile from one server. Suppose Riis the peer connected
to the server, and Tiis the subtree rooted at Ri,andLi
is the set of leaves of Ti. To get a quick downloading
time, FOX fills the available outgoing link at the server, in
which each Rirequests a different block from the server,
and each leaf only sends blocks to leaves which reciprocate
the gesture. Peers in FOX have two properties. The first
property is the give-and-take symmetry, which means each
of peer’s outgoing edges has a corresponding incoming edge,
instead of allowing peers to arbitrarily match up. The second
property is that the connections are stable. Based on these
two properties, peers’ misbehavior, such as receiving but not
sending, can be detected. In addition, FOX can deal with the
last-block collapse, in which the finished nodes quit the system
leaving the remaining nodes without any help. Fox uses a
cryptographic method, in which the final blocks are encrypted
by the internal nodes before sending to the leaves. Then, the
leaves use their own keys to encrypt the encrypted blocks.
So both nodes and leaves need others’ keys before decrypting
the encrypted blocks and completion of download. By using
this approach, both internal nodes and leaves hold information
that the others need to complete their download. So FOX can
make all nodes complete their download at the same time and
avoid the last-block collapse.
In P2P systems, users often decide which block to download
without considering whether the block has already been sent
out by the server at least once or not. This causes subop-
timal distribution decisions. Gkantsidis et al. [10] proposed
Avalanche using networking coding technique. Like BitTor-
rent, nodes join the system by contacting a centralized server
which provides a random subset of other users. Nodes also
employ rarest first algorithm to decide which block to transfer.
In Avalanche, there are two incentive mechanisms. The first
one is giving priority to exchange over free uploading to
other nodes. The second one is that nodes do not upload
to users unless they have also received sufficient content in
return, which is similar to the TFT policy. The principle of
network coding is to allow nodes to encode packets. When a
client wants to send a packet to others, it will make a linear
combination of all the available information and then send
it out. At the receiver’s side, the original information will
be reconstructed after receiving enough linear combination
of packets. With this method, the block propagation becomes
more efficient and the system is more robust under extreme
situations like the sudden departure of nodes.
The characteristics of the above BitTorrent-like systems are
summarized in Table I. As can be seen from the table, the
above three systems not only share some mechanisms in com-
mon with BitTorrent, but also have their own characteristics.
Slurpie is very close to BitTorrent in its operation. A major
difference is that it provides bandwidth estimation technique
which enables it to scale well as the group size increases. FOX
proposes a fair protocol to provide an optimal downloading
time. While Avalanche is very similar to BitTorrent, it uses
network coding technique to make the scheduling of the
content propagation easier.
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

144 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
TAB L E I
BITTORRENT-LIKE SYSTEMS
BitTorrent-like system Special Characteristics Common characteristics with BT References
Slurpie Bandwidth estimation technique. The number of neigh-
bors is updated according to the available bandwidth
Central server. Mesh formed
among peers.
[12]
FOX Peer connections are based on give-and-take symmetry.
The connections are stable.
TFT policy [11]
Avalanche Each piece of the shared file is encoded by the nodes
using network coding.
Central server. Rarest first policy.
Incentive mechanism.
[10]
III. PERFORMANCE OF BITTORRENT
The popularity of BitTorrent has spurred a large number
of studies of its performance. Some measure its behavior on
real torrents; others use the simulation to observe its perfor-
mance while others use mathematical analysis-based models
to evaluate its performance. In this section, we first organize
the papers based on the performance evaluation method used.
The main purpose of organizing this section according to the
performance evaluation methods used is to group together
the papers using similar approaches to evaluation. We do
notice that each evaluation technique lends itself to examining
certain aspects of BitTorrent performance better than the other
techniques. We highlight the advantages and shortcomings of
each approach in the respective section. In the final subsection
we concentrate more on reviewing the techniques from the
perspective of the issues addressed and the different aspects
of performance studied.
A. Measurement
Several measurement-based studies of BitTorrent perfor-
mance have been presented in the literature. Most of the
measurement-based experiments lasted for several months.
Generally, the following five methods were used: (a) ana-
lyzing the tracker logs obtained from the trackers, (b) using
scripts to gather information from torrent websites and directly
from peers, (c) analyzing packet traces collected at Internet
access link, (d) joining an ongoing torrent with a modified
client specially instrumented to collect event logs, and (e)
conducting experiments on network testbeds like PlanetLab
or user-constructed networks of PCs. Observing the tracker
log can enable us to get the information on the global view
of BitTorrent performance; whereas using a modified client
enable us to observe the individual behavior of peers.
Izal et al. [18] evaluated the performance of BitTorrent
mechanisms based on five months of measurement data
collected from BitTorrent peers. This was one of the first
comprehensive evaluation of BitTorrent using measurement.
First, they used the tracker log of Linux RedHat 9 distribution
for their evaluation. They found that:
•BitTorrent clients are altruistic.
•Seeds contribute more than twice the data uploaded by
leechers.
•The proportion of seeds is higher than 20%.
Then, they used an instrumented client to join the torrent to
observe the individual behavior of a peer. They observed that:
•After obtaining the first few pieces, peers begin trading
pieces soon, which indicates that the rarest first policy
works well.
•The uploading and downloading rates are correlated,
which means that the TFT policy works well.
We note that the perspective presented in this paper is based
on the detailed analysis of a single torrent, which may be
its limitation in that the perspective is limited. However
the observations are still relevant and corroborated by other
studies reviewed in this section later.
Pouwelse et al. [6] presented a measurement based study
of BitTorrent’s availability, integrity, flash crowd handling, and
download performance based on measurement data obtained
from June 2003 to March 2004. They also measured Suprnova,
one of the most popular websites for BitTorrent clients to find
files. Their experiment consisted of two parts. They monitored
the global BitTorrent components using three scripts: a Mirror
script for measuring the Suprnova mirrors’ availability and
response time, a HTML script for gathering and parsing
Suprnova mirrors’ HTML pages and downloading new files
with the extension of .torrent, and a Tracker script used to
parse the .torrent files for new trackers and check the trackers’
status. They also observed the actual peers in the system using
scripts: the Hunt script was used to select a file to follow and
initiate a measurement of all the peers who are downloading
that file, the Getpeer script was used to obtain the IP addresses
of peers downloading the file from the tracker, the Peering
script for measuring the download progress and uptime of
peers in parallel. They found that:
•The number of users is related with the number of seeds
in BitTorrent.
•Most users often stay in the system to be seeds for several
hours after finishing their downloading.
•BitTorrent can deal with flashcrowds efficiently.
•Fewer number of seeds does not mean that the lifetime
of the file is short. Even files with only one seed can still
have a long content lifetime.
The authors concentrated mainly on the user behavior and the
content distributed by BitTorrent. This gives a good macro
view of BitTorrent behavior. However this measurement does
not directly shed light on the core mechanisms in BitTorrent
and their effect on its performance.
Legout et al. [22] conducted a detailed measurement study
aimed at understanding the behavior of the rarest first al-
gorithm and the choking algorithm on real torrents. They
instrumented a BitTorrent client and observed its performance
in several torrents. They observed the log of messages sent
or received, the log of each state change, the log of the
rate estimation, and the log of important events such as end
game mode. Their results showed that (1) the rarest first
algorithm prevents the reappearance of rare pieces and avoids
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 145
TAB L E I I
SUMMARY OF MEASUREMENT-BASED STUDIES OF BITTORRENT PERFORMANCE
Measurement Methods Results References
Observing the tracker log BitTorrent clients are altruistic. The rarest first policy works well. The
TFT policy works well
[18]
BitTorrent can support large files. The incentive mechanisms are effec-
tive
[19]
TFT is imperfect at preventing free-riding [20]
The distribution of performance among peers is roughly uniform. TFT
is the main factor that affects the performance
[21]
Using Mirror script, HTML script, Tracker
script, Hunt script, Getpeer script, and Peering
script
BitTorrent can deal with flash crowds effectively [6]
Analyzing packet traces collected at Internet
access link of a residential university network,
and Tracker Logs
BitTorrent is locality-unaware, causing the same content to be down-
loaded into the same locality (defined around an ISP) multiple times.
Significant overlap in lifetimes of peers within the same locality.
Incentives to peers to stay after download completes to serve others
has significant impact on performance
[8]
Using a client on real torrent The rarest first algorithm prevents the reappearance of rare pieces and
avoids the last pieces problem. The choke algorithm is fair and robust.
[22]
Peers behind a firewall have low bandwidth usage. Peers behind a
firewall often disconnect early
[23]
Using a modified version of BT on PlanetLab The choking algorithm can cluster peers with similar bandwidth, reward
contributing peers, and make peers have a high upload utilization
[24]
The initial stage is not predictive of the overall performance. The
unchoked network is scale-free. There is no clustering except the initial
stage.
[25]
Doing experiments on a testbed of 96 PCs Random neighbor selection causes high operating cost [26]
the last pieces problem; (2) the replacement of block level
TFT with a bit level TFT solution does not necessarily yield
better fairness; (3) the choke algorithm is fair and robust.
Their recommendation is that the rarest first strategy already
achieves excellent performance and there is no compelling
need for more complex strategies. They do observe some scope
for improvement in the choke algorithm. While this study
sheds light on the important properties of the choke algorithm,
however their measurement is from a single-peer view point
which cannot evaluate the dynamics of BitTorrent. Also, the
reasons behind the observed properties of the choke algorithm
are unclear.
Bellissimo et al. [19] collected statistics from two trackers
to conduct their measurement-based evaluation. The statistics
included the user’s random ID, how long the user had been
connected to the torrent, and the number of bytes uploaded,
downloaded, and remaining. Their results showed that (1)
BitTorrent can serve large files effectively; (2) the incentive
mechanisms are effective.
In [23], Skevik et al. focused on the effect of the firewalls
and the time the peers have stayed in the system. They
monitored the torrents of RedHat and Mandrake Linux for
several months by executing a modified client and a crawler.
They observed that peers behind a firewall have low bandwidth
usage and they are more likely to leave the system early
because of their poor download performance.
Karagiannis et al. [8] quantified the influence of BitTorrent
on ISPs using two types of real measurement data, the
BitTorrent tracker log for RedHat v9.0 and the payload packet
traces collected using a monitor installed on the access link
of the network of a residential university in USA with 20,000
users. While the paper considers the more general issues of
peer-assisted content delivery and its impact on ISPs, we
are specifically concentrating on their measurement based
evaluation and observations about BitTorrent. They indicated
that BitTorrent is locality-unaware which severely increases
the ISPs’ cost, typically resulting in the same content being
downloaded into the same ISP multiple times from external
peers. They also found that users requesting the same content
have a 30%-70% overlap in their lifetimes, consequently
helping each other in the download. An additional 20%-40%
improvement in terms of cross-ISP downloaded content can
be effected by giving incentives for users to stay around after
they complete the download.
Andrade et al. [20] focused on measuring the cooperation in
BitTorrent. In their measurement, they defined three metrics:
(1) free-riding ratio, which is the percentage of free-rider
peers, (2) seeding ratio, which is the percentage of seeds in
a torrent, and (3) sharing ratio, which is the total uploaded
data divided by the downloaded data. They evaluated the
metrics by collecting logs in BitTorrent communities such
as bt.etree.org, and piratebay.org. Then, they measured the
downloading time for free-riders and non free-riders, and the
amount of free-riding and low-sharing behavior in BitTorrent
communities: etree and easytree. Their results demonstrate
that BitTorrent’s TFT protocol does discourage free-riding
successfully. However, if there are a large number of seeds
in the torrent, the TFT mechanism may not work effectively
to prevent free-riding.
Legout et al. [24] conducted experiments on private torrents
and collected data from peers in a controlled environment.
They focused on the choking algorithm in BitTorrent to
observe peers’ individual behavior during the downloading
process. Their experiments were performed on PlanetLab.
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

146 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
They collected logs of sent or received messages, and state
changes using a modified version of BitTorrent. Their results
reveal that the choking algorithm can cluster peers with similar
bandwidth, reward contributing peers efficiently, and make
peers have a high uploading utilization.
Dale et al. [25] looked at the topologies of different kinds
of networks that can be clearly identified within a BitTor-
rent swarm. In particular they investigated the evolution of
four kinds of networks that can be used to characterize a
swarm: (1) Connection Network: the network of neighbors
that each peer maintains. (2) Interest Network: the network
defined by the interest that peers maintain in other peers.
(3) Unchoked Network: the network that is formed by the
peers and the neighbors that are unchoked by each peer. (4)
Download Network: the network that is formed linking peers
to other peers from which they are downloading. Among
these networks, all networks are directed graphs except for
the connection network which can be considered undirected.
In their experiments, they used a modified BitTornado [27]
client running on more than 400 nodes of PlanetLab that form
the swarm. They found that:
•The initial stage is not predictive of the overall perfor-
mance.
•The unchoked network is scale-free.
•There is no evidence of clustering in their experiment
except in the initial stage.
Rasti et al. [21] pointed out that several of the measurement
studies of BitTorrent were just based on observing the behavior
of a few instrumented peers. This may not necessarily provide
a representative view of the overall BitTorrent performance.
Instead, a distribution of the performance indices among the
peers provides a broader view of BitTorrent performance.
Furthermore, the effect of different peer-level and group-level
properties on the performance needs to be identified. To verify
their theory, they analyzed a BitTorrent tracker log to estimate
the distribution among peers in a torrent and identify the effect
of the peer-level and group-level properties on the perfor-
mance. They found: (1) the distribution of performance among
peers is roughly uniform; (2) the TFT mechanism is the main
factor that affects the performance. They could not observe
any direct relationship between peer-level performance and
main peer and group-level properties. Furthermore, the average
upload rate of individual peers had the highest correlation with
its observed performance [21].
Lei et al. [26] did an experiment to test the performance
of BitTorrent’s neighbor selection. Their experiment includes
96 PCs, each of the same configuration. They used BitComet
as the BitTorrent client for their experiments. They observed
that the random neighbor selection in BitTorrent does not take
into account the communication cost, which will result in low
transmission rate and high operating cost.
We summarize the various observations on BitTorrent per-
formance obtained through measurements in Table II.
B. Simulation
Evaluating the impact of various factors on the overall
performance is very difficult using measurement-based tech-
niques. Simulation based studies on the other hand provide
much greater flexibility to vary the various mechanisms used
in BitTorrent and observe the impact through extensive exper-
imentation. Compared with the measurement-based approach,
the simulation-based approach has two advantages. One ad-
vantage is the greater flexibility in controlling the various
configuration parameters of BitTorrent mechanisms. Another
advantage is that it allows us to study the impact of variations
in a particular mechanism while keeping the rest fixed, which
is very difficult to achieve in measurement-based experiments.
Simulation-based approaches are especially useful in topology
related studies since a crawler does not yield information
about a peer’s neighbors in BitTorrent because of the lack of
distributed mechanisms for peer discovery or lookup. Further-
more a peer’s connectivity information is not shared with the
tracker and hence tracker logs do not yield overlay information
[31].
Many of the simulation based studies make simplifying
assumptions for the sake of ease of simulation. Sometimes,
these assumptions may not necessarily be realistic. Very often
the underlying network topology is simplified and packet-level
dynamics are not fully represented. Most of the simulators
concentrate only on representing the behavior at the overlay
network level. Thus the impact of the underlying physical
network is not clearly reflected in the simulations. Also many
simulations consider a smaller set of peers in a swarn than is
normally encountered in practice. The results from simulation
studies are summarized in Table III.
Felber et al. [28] presented a simulation based study of
the effect of different peer and piece selection strategies on
the performance of BitTorrent. In their peer selection strategy,
they assumed that a peer has a global view of the state of
all the other peers and thus selected the most suitable peer
to exchange pieces with based on the peer selection strategy.
They considered several peer selection strategies including
random, and several variations based on the number of pieces
possessed/missing in the other peers. Their results demonstrate
that there is no clear winner among all the strategies in
all scenarios. Similarly, they considered two different piece
selection strategies: random and rarest. Among these the rarest
strategy seems to deliver the most consistent performance
across different peer selection strategies.
Bharambe et al. [29] evaluated the impact of BitTorrent’s
core mechanisms on overall system performance under a
variety of flash-crowd workloads using a simulator, which
can model not only peer’s activity such as joins, departures,
and block exchanges, but also BitTorrent’s mechanisms like
TFT and rarest first. To make the model simple, they did not
model network propagation, the packet-level dynamics of TCP
connections and the shared bottleneck links in the interior of
the network. They focused on the link utilization, the fairness
and the optimality to quantify the effectiveness of BitTorrent.
They made their simulator named BTSim publicly available
[34]. From the simulation results they found that:
•BitTorrent is robust and scalable, and ensures high uplink
bandwidth utilization.
•The TFT policy fails to prevent the unfairness in BitTor-
rent.
•The rarest first policy is critical in ensuring that the newly
joined peers have something to exchange with others.
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 147
TABLE III
SUMMARY OF SIMULATION-BASE D STUDIES OF BITTORRENT PERFORMANCE
Simulation Features, Assumptions and Shortcomings Results References
Event-driven simulator with resolution of a millisecond.
Peers have global view for peer selection
No appreciable impact of peer selection strategies. Rarest first
piece selection strategy gives balanced performance
[28]
The event-driven simulator models both peer’s activ-
ity and BT’s mechanisms. Network propagation, the
packet-level dynamics of TCP connection and the
shared bottleneck links are not modeled
BT is robust and scalable at ensuring high uplink bandwidth
utilization. The TFT policy fails to prevent unfairness in Bit-
Torrent. The rarest first policy is critical to ensure the new joined
peers have something to exchange with others.
[29]
Simple simulator contains one seed serving a file com-
posed of npieces. Network latency, connection costs
or complex topologies not modeled.
By selectively uploading to its poorest neighbor, the peer
ensures that the data remain in the network for the longest
period, thus reducing its own burden.
[30]
The simulator is developed in MATLAB The authors
simulated the tracker protocol in mainline client 4.0.2.
The overlay of BT is not a random graph. A large number of
NATed peers decreases the robustness of the overlay to attacks
significantly.
[31]
The simulator runs in rounds and each round lasts for
10 seconds A leecher sorts peers based on the data they
offer The bottlenecks is downlinks/uplinks. The authors
did not model any delay.
The peer set size (PS) and the percentage of outgoing con-
nections (OC) have a significant impact on the BitTorrent’s
performance. Decreasing OC is more efficient than increasing
PS. The choice of PS and OC can have an impact on the overlay
structure.
[32]
The set of nodes involved in BT session is fixed.
Endgame mode is not simulated. The neighbor set of
each BT node is static. Delays are not simulated except
network packet transmission delay.
The current BT is far from the optimal solution that minimize
the end-to-end delay of distributing a file from a source to
multiple receivers. The peer selection strategy may lead to
mismatched peering and make BT deviate more from optimum.
[33]
The simulations were restricted to a peer set no larger than 15
peers, while the real implementations can have up to 80. This
might influence some of the results since the accuracy of the
piece selection strategy is affected by the peer set size.
Adar [30] demonstrated through simple simulations that a
peer possessing most of the file pieces can significantly reduce
its effort in servicing other peers while at the same time
improving network performance. They simulated the effect of
bit welfare. There is only one seed serving a file composed
of npieces in the simulator and three seed strategies are
considered: Random, in which the seed chooses the peer to
give piece randomly, Preferred, in which a peer is chosen to
receive pieces from a preferred subset, Poorest-first, in which
the peer who has the fewest pieces is chosen. Based on their
experiments, they found that the Poorest-first runs better than
other algorithms. The experimental results showed that by
selectively uploading to its poorest neighbor, the peer ensures
that the data remain in the network for the longest period, thus
reducing its own burden.
Urvoy-Keller et al. [32] adopted a simulation approach
to evaluate the impact of the overlay topology parameters
on the BitTorrent performance. Their results show that the
two parameters, the peer set size (PS) and the percentage of
outgoing connections (OC), have a significant impact on the
BitTorrent’s performance. In addition, decreasing OC is more
efficient than increasing PS. Furthermore, the overlay structure
can be affected by the choice of PS and OC.
Al Hamra et al. [31] followed up on the work by Urvoy-
Keller et al. [32]. They specifically concentrated on evaluat-
ing the properties of the distribution overlay of BitTorrent.
They presented an in-depth study of the overlay topologies
in BitTorrent by conducting simulation experiments using a
simulator developed in MATLAB. They analyzed the relation
between the overlay properties and the BitTorrent performance
considering the following four parameters: average peer set
size, the time for a peer to reach its maximum peer set size,
the diameter of the overlay, the robustness of the overlay to
attacks and high churn rate. They pointed out that a large peer
set size can make BitTorrent more efficient. More importantly,
they showed for the first time that the overlay in BitTorrent is
not a random graph, and connectivity of a peer to its neighbors
depends on its arriving order. In addition, they showed that
a large number of NATed peers (peers behind a NAT or a
firewall) significantly decreases the robustness of the overlay
to attacks.
Wu et al. [33] were specifically interested in answering the
question of how close BitTorrent is to the optimal solution.
They proposed a centrally scheduled file distribution (CSFD)
protocol that can minimize the total elapsed time. In addition,
they compared the performance of BitTorrent and CSFD. They
found that the current BitTorrent is far from the optimal
solution that minimizes the end-to-end delay of distributing
afile from a source to multiple receivers. Moreover, the
peer selection strategy of BitTorrent may lead to mismatched
peering, and make BitTorrent deviate from the optimum.
C. Analytical Modeling
Several analytical models of BitTorrent-like file sharing
system have been proposed in the literature (See Table IV).
Existing analytical studies can be classified into two cate-
gories: homogenous, where all peers have the same upload
and downloading rate; and heterogeneous, where peers have
different uploading and downloading rates. Analytical models
are clearly able to reflect the effects of different parameters
on BitTorrent performance. They permit efficient and detailed
exploration of the parameter space to evaluate the effect of
variations of not just a single parameter, but also the combined
effect of variations of several parameters. However, many of
these models are based on sometimes unrealistic assumptions
like peers having global information about the state of all
peers, simplifying assumptions on the underlying network
topology, and on the arrivals and departures of peers. Most of
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

148 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
TAB L E I V
SUMMARY OF ANALYSIS -BASED STUDIES OF BITTORRENT PERFORMANCE
Focus of Analys is Modeling
Approach
Results References
Service capacity and fairness Homogenous
Branching process,
Markovian
P2P systems achieve favorable scaling in terms
of average download delay with increasing load.
ε-peerwise fairness is a practical approach, for
example as used in BitTorrent.
[35]
Scalability and incentive mechanisms Homogenous
Fluid-Flow
The system scales well with the number of peers.
The incentive mechanisms are efficient. Optimistic
unchoking may encourage free-riders
[3]
Single-Torrent system, file availability
Multi-torrent system performance
Homogenous
Fluid-flow
BitTorrent is good at dealing with “flash crowds.”
Seed departures and decreasing peer arrivals causes
file to become unavailable soon. Peers with higher
downloading capacity do not upload as much as
slower peers. Inter-torrent collaboration is effective
to improve performance
[36]
The distribution of the individual chunk Homogenous The modified routing policies perform better [37]
Evaluation of file distribution time
Application-level differentiated service
model
Heterogeneous
Fluid-Flow
Obtains an explicit expression for the minimum
achievable time to distribute a file.
[38]
Influences of free-riding Heterogenous
Fluid-flow
BitTorrent mechanism can prevent free-riding suc-
cessfully
[39]
The distribution of the peers’ download
state, file availability and system death
Homogenous
Fluid-flow,
Markovian
Peer distribution in terms of their current download
state follows a U-shaped curve, with most peers in
either just starting the download or possessing the
entire file. The seeds’ departure rate and leechers’
abort rate can influence the peer distribution. The
TFT strategy cannot improve the file availability or
prevent system death.
[40]
The download behavior and incentive
mechanism
Homogenous
Queueing Model
First attempt at a queueing model for BitTorrent
protocol
[41]
Altruism and Incentive Mechanisms. Fo-
cus on reciprocation and upload band-
width sharing
Homogenous
Probabilistic
model
Altruistic contribution by high-capacity peers seem
to cause the most benefit in download performance.
TFT policy is not the major cause for good perfor-
mance
[42]
Effect of Bandwidth Heterogeneity on
download performance
Heterogeneous
Fluid-flow
Heterogeneous bandwidth can have a positive ef-
fect on content propagation
[43]
Two-class problem. Focus on Service
differentiation among the classes. Band-
width diversity problem
Heterogeneous
Fluid-flow
System of differential equations admits a unique
stable equilibrium point. Whether the system of
differential equations has a stable state or not
depends on the initial conditions.
[44]
File downloading time File availability
Both Single class and Multiclass
Heterogeneous
Fluid-flow
Stochastic differential equation approach, yields
closed form solutions for several performance mea-
sures including average download time, system
throughput, number of peers and seeds. Sensitivity
analysis of performance measures.
[45]
File dissemination in P2P networks, Min-
imum file distribution time
Heterogeneous de-
terministic
Distribution time of any natural random strategy as
adopted by BitTorrent is proportional to the optimal
time
[46]
Heterogeneous peers and performance
using simple models using balance equa-
tions.
Heterogeneous
deterministic flow
balance equations
Simple model yields accurate results, verified with
simulation. Token based scheme, which can yield
better performance for low-bandwidth peers by
giving higher weightage to their upload contribu-
tion.
[47]
the analytical models rely on a fluid-flow approach to charac-
terize the system behavior. In particular, the models represent
the evolution of the system by tracking the number of peers,
X(t), and the number of seeds, Y(t), either as a function of
time, or in steady-state. The evolution is represented using
(stochastic) differential equations involving various system
parameters. Typical parameters that characterize the system
evolution include:
•λ, the arrival rate of new requests. The interarrival time
is usually assumed to be exponentially distributed.
•μ, the upload bandwidth of a peer.
•c, the downloading bandwidth of a peer.
•θ, the rate at which peers abort their download
•γ, the rate at which seeds leave the system, and
•η, the effectiveness of file sharing.
Some models keep track of the different stages of downloading
that a peer may exist at any point of time. This typically
involves dividing the peers, X(t), into different subsets,
X1(t),X
2(t),etc. each representing the different stages of
download. However, this increases the number of simultaneous
equations that need to be solved and hence the complexity of
the model.
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 149
Yang et al. [35] modeled and analyzed P2P systems in
terms of their “service capacity” for both transient and steady-
state regimes. They considered an abstract model of P2P
file sharing system, and focused on peers that are sharing a
single file. They concentrated on the file download process and
assumed that each peer knows the address and file availability
at other peers. Furthermore, they assumed that no peer leaves
the system before finishing their downloading. They first
presented transient analysis of the P2P system using both
deterministic and branching process approaches to determine
the average download delay. They demonstrated that P2P
systems achieve favorable scaling in terms of the average
download delay with increasing load. They also show that
parallel uploads from a peer, where the upload bandwidth is
used by a peer to simultaneously upload to multiple peers,
is effective even when peers are uncooperative. Then they
modeled the stationary regime using a Markovian approach.
Furthermore, they studied the fairness issue in P2P systems by
considering different notions of fairness for service allocation
in the stationary regime. Achieving global proportional fair-
ness and pairwise proportional fairness is not easy for robust
implementation. They then proposed an ε-peerwise fairness
which is better suited for a dynamic environment with peers
joining and leaving the system. This is the strategy used in
BitTorrent with its choking algorithms.
Qiu et al. [3] presented an analytical model of BitTorrent
based on a fluid flow model. Their work was influenced in
part by the earlier work of Yang et al. [35]. In developing their
model, they made several assumptions: downloaders become
seeds with a probability that is determined by parameters like
the number of peers; downloaders can abort their downloading
after a certain time that is exponentially distributed, and
each seed stays in the system for a random time which
is also exponentially distributed. They developed a simple
deterministic fluid-flow model to describe the evolution of
the number of leechers and the number of seeds, in terms
of the arrival rates, the upload and download bandwidths,
the abort rate of downloaders, the rate at which seeds leave
the system and the effectiveness of file sharing. They used
this model and studied the steady-state performance, with
special attention to the protocol’s scalability and efficiency.
They found that the average downloading time is not related
to the peer arrival rate, and hence the system scales very well
with the number of peers. Their model clearly demonstrated
that the average downloading time decreases with efficient file
sharing and increases with increasing seed departure rate, as
expected. Also changes in the uploading and downloading
bandwidth impact the downloading delay to some extent.
They then presented a game-theoretic analysis of both the
choke and rarest first algorithms. They show that the incentive
mechanisms in BitTorrent are efficient. They also show clearly
that the optimistic unchoking might lead a free-rider to get
upto 20% (In general, 1/(nu+1)%, where nuis the number
of regular unchokes) of the maximum downloading rate.
They also presented experimental results from simulations to
validate the results from their models. The assumption about
global knowledge of all peers for peer selection is a major
limitation of this approach.
Guo et al. [36] followed up the idea of [3]. First they
analyzed the file downloading trace files obtained from two
dedicated tracker sites that host multiple torrents. They also
analyzed BitTorrent metafile downloading traces that were
collected from a large commercial server farm hosted by a
major ISP and a large group of users connected to the Internet.
From the trace analysis, they concluded that the peer arrival
rate to a torrent decreases exponentially. They studied the
torrent lifespan, the duration from birth until the file becomes
unavailable. They derived the expressions for the torrent
lifespan and the downloading failure ratio, the ratio of peers
that do not complete the download to the total peer population
for a torrent. They extended the fluid-flow model in [3] using
the model for the exponentially decreasing peer arrival rate
to construct the torrent evolution model. They found that
both peers and seeds increase exponentially initially, but then
decrease exponentially at a slow rate. The seed departures
result in the torrent being unable to keep up with the peer
service demand, causing the torrent to eventually die. The
random arrival and departures of peers and seeds causes the
performance to fluctuate significantly over the torrent lifetime,
especially for small torrents. Furthermore, they demonstrated
that the biased service policy of seeds towards peers with
higher download bandwidth might result in these peers not
contributing their fair share of upload capacity to the torrent.
They also proposed a graph based multi-torrent for analyzing
inter-torrent collaborations. Their analysis demonstrated the
feasibility of using multi-torrent collaboration effectively to
improve performance.
In [37], Arthur et al. proposed a BitTorrent-like model
using which they analyzed the data disseminating protocols
of BitTorrent and related P2P networks. In this model, they
ignored the TFT strategy and assumed that all nodes have the
same bandwidth. In particular, they modeled the routing of
data blocks on a directed graph over discrete time steps. They
also analyzed some of the multiple network topologies and
routing algorithms.
Kumar et al. [38] presented an analytical model of peer-
assisted file distribution time in P2P networks using a fluid-
flow based model. Their aim is to study the advantages of
a peer-assisted file distribution system vis-`a-vis a traditional
client-server distribution system, specifically the scalability
and efficiency of the new approach. In particular, they were
interested in finding the minimum distribution time for the
file to be distributed from the seeds to all the leechers in a
general heterogeneous peer-assisted file distribution system.
In their model, they made two assumptions: The bandwidth
bottlenecks are in the uploading and downloading link rates
instead of in the Internet core, and every node participates in
the file sharing until it completes the file download. Using
this model they obtained an explicit expression for the min-
imum distribution time. This expression explicitly shows the
relationship of the distribution time to the file size, the seeds’
uploading rate, and the leechers’ uploading and downloading
rate. They also point out that although their approach makes
an unrealistic assumption about the flow of bits through a
node compared to the chunk-based models that were proposed
earlier, the resulting expressions provide a very good (lower-
bound) approximation for the file distribution time. However,
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

150 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
it is important to note that peer heterogeneity is better handled
by their approach.
In [39], Yu et al. focused on analyzing the influence of free-
riding in BitTorrent System. They extended the model in [3]
with two classes of peers: nonfree-riders and free-riders. Their
results show that BitTorrent’s mechanism can guard against
free-riding successfully. In addition, they discussed the system
death induced by free-riding and pointed out that retaining a
copy of each piece in the system can prevent system death.
Tian et al. [40] extended the fluid-flow model in [3] to
study the distribution of the peers in different states of the
downloading completion. Towards this end, they explicitly
represent the peers as belonging to one of several states,
S0through SN,where1/N is the level of granularity at
which they view the current download progress. The peer
state behavior is now viewed as a continuous time Markov
chain. From this they derived the distribution of the peers in
each state of the download. The distribution is observed to
be a U-shaped curve. From the model, they found that seeds’
departure rate and leechers’ abort rate can influence the peer
distribution. Furthermore, the TFT strategy cannot improve
the file availability or prevent the system death caused by a
sudden departure of a finished peer.
Saraswat et al. [41] presented a queuing model for BitTor-
rent using the Java Modeling Tools [48]. In their model, each
BitTorrent client is modeled as a load dependent station. The
stations’ service request time corresponds to the time taken to
download a piece, the service time correspond to the inverse
of the downloading bandwidth, and the number of jobs in
the queue represents the total number of pieces downloaded.
They set the service time dependent upon the number of jobs
in the station to ensure that the incentive mechanisms work
well. They focused on the download behavior and incentive
mechanisms.
Piatek et al. [42] presented a model for the altruism in
BitTorrent. They made the following assumptions: (1) The
distribution of a typical swarm may not be identical; (2) The
active set sizing is uniform; (3) There is no steady state;
(4) Swarm performance is limited by upload bandwidth. The
authors built the model based on the expressions for the
probability of TFT reciprocation, expected downloading rate,
and expected upload rate. In the model, they considered two
definitions of altruism. It can be defined as the difference
between expected upload rate and the download rate. Another
definition is any upload contribution that can be withdrawn
without loss in download performance. Their result shows that
the TFT policy is not robust, and the altruistic contribution
plays an important role in BitTorrent’s performance.
The first heterogeneous fluid model of BitTorrent-like file
sharing system was proposed by Lo Piccolo et al. in [43].
To assess the effect of different access link capacities on
BitTorrent-like file sharing system, they developed a fluid
model with access links of two different capacity classes
by extending the fluid-flow model presented in [3]. To keep
the heterogeneity simple, they only considered links of two
different link rates, viz. a high speed downloader and a low
speed downloader. Furthermore, they assumed that the links
are symmetric, i.e., the upstream rate is the same as the
downstream rate. They analyzed the effects of bandwidth
heterogeneity on file transfer dynamics and content diffusion
process in detail. In particular, they compared the performance
of heterogeneous networks and the equivalent homogeneous
networks under different conditions of equivalence. Their
results show that heterogeneous bandwidth can have a positive
effect on content propagation among peers if appropriate
criteria were chosen.
Cl´evenot-Perronnin et al. [44] argued that the model in [3]
considers only homogeneous peers, but in practice, peers have
diverse bandwidth characteristics, such as Ethernet access,
dial-up modem access and broadband access. This limits
the model’s applicability. Therefore, they presented a general
multiclass model for heterogeneous peers with different ac-
cess bandwidth and multiple differentiated service classes. In
particular they considered in detail a system with two classes
of peers, distinguished by their different upload and download
speeds. In their model, they allow an uploader (either a seed
or peer) to statically allocate its upload bandwidth to the
two classes of peers. They used this model to analyze two
problems: service differentiation and bandwidth diversity. For
the service differentiation problem, they described that the
system of differential equations admits a unique stable equi-
librium. For the bandwidth diversity problem, they indicated
that whether the system of differential equations has a stable
state or not depend on the initial conditions.
Fan et al. [45] extended the model in [3] to obtain a
fluid-flow model based on the stochastic differential equation
approach. In this model, they divided peers into three types:
leechers that have a few chunks, leechers that have most
chunks, and seeders. They define the connection probability
ρ(≤1) as the probability that a peer maintains connectivity
with another peer in the torrent. They analyzed the file
downloading time and file availability in BT-like systems
with this model. They obtained closed-form solutions for the
average number of seeds and leechers, the average down-
loading time and the steady state throughput of the system.
The closed-form solutions show the effect of various system
parameters including the peers’ arrival rate, seeds’ departure
rate, connection probability and transmission bandwidth on the
performance measures explicitly. They also carried out sensi-
tivity analysis of the performance metrics with respect to the
different parameters. They used a discrete-event simulator to
validate the results obtained from analysis. They then extended
the model to account for peers situated behind firewalls and
discussed the impact on the system performance. Their results
clearly show that those peers not behind firewalls play an
important role in determining the overall system performance.
They also discussed the effect of different chunk selection
algorithms at the peers on the file availability within the
system. Through mathematical analysis they showed that the
rarest first policy indeed is the best to increase file availability.
In practice this is true for low connection probability among
peers. However when ρis large, this may cause a reduction in
file availability. To improve file availability, they proposed a
file availability enhancement algorithm which randomizes the
selection of the chunk to download, but still giving preference
to the rarest chunk with higher probability. They showed
that this enhancement improves the file availability in all
circumstances.
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 151
Mundinger et al. [46] studied the file dissemination in P2P
networks using the ‘uplink sharing’ approach, which is a
variation of the classic broadcasting problem. In their model,
nodes have different upload capacities. They assumed that
the available upload capacity is shared equally amongst the
connections, and uploads are not interrupted until complete,
which means the rate is always positive. They showed that
a simple and natural random strategy adopted by BitTorrent
results in a distribution time which is proportional to the
optimal time achievable.
Liao et al. [47] considered heterogeneity in a BitTorrent
system by distinguishing users into two classes, viz. high
bandwidth users who have high upload-link capacity, and
low bandwidth users with low upload-link capacity. They
proposed a mathematical model to predict the average file
download delay for the two classes of clients. This model
represents the system behavior in steady-state by matching
the amount of data uploaded and downloaded by the peers.
In addition, they designed a token-based TFT scheme. In
this scheme tokens are used as a means for deciding how
a peer uploads data to its neighbor. Tokens are earned by a
node by uploading to its neighbor. A node keeps track of
the number of tokens for its neighbors and makes its upload
decisions based on the tokens. By using a biased means for
a low-bandwidth peer to earn more for uploading to a high-
bandwidth peer, the authors design a system which ensures
better download performance for low-bandwidth peers, albeit
with a higher upload contribution by the high-bandwidth peers.
They validate their model by comparing the results with
simulation results for several experiments conducted using the
BTSim simulator provided by [34] for different scenarios. The
results show that the proposed mathematical model is quite
accurate in predicting the file download performance.
D. Summary and Comparison of Performance Studies
In this section, we examine the performance studies from
the perspective of the different aspects of the BitTorrent
protocol that were studied. We summarize our findings in Ta-
ble V, in particular highlighting how the different aspects were
studied using the different performance evaluation methods,
with more or less the same conclusions.
All the three performance evaluation methods have focused
on evaluating the piece exchange mechanisms in BitTorrent.
The general findings from the different approaches are in
agreement with each other. In particular, it is found that the
TFT mechanism is very useful in ensuring co-operation among
the peers. Furthermore, the general observation is that the
presence of altruistic peers encourages free-riding. Similarly,
the rarest first policy has been found by all methods to be very
effective in increasing download performance and decreasing
the file download time.
Topology related studies are feasible only through measure-
ment and simulation, since these methods explicitly reflect the
effect of the topology on performance. In particular, the topol-
ogy’s effect is direct in measurement. In simulation studies, the
topology is usually abstracted at the overlay topology level.
Analytical modeling does not explicitly track the topology,
but its influence may be reflected through system parameters
like the connection probability ρin [45]. References [25]
and [24] studied the BitTorrent topology using measurement-
based studied, while [31] and [32] studied the same topic by
simulation. They observed the overlay structure and evaluated
the parameters which can impact the overlay topology. In
addition, BitTorrent’s neighbor selection and the impact of
BitTorrent on ISP traffic were also measured by [8] and [26].
They show that the random neighbor selection and BitTorrent’s
locality-unawareness have a significant impact on the overall
performance.
IV. IMPROVEMENTS TO BITTORRENT MECHANISMS
The phenomenal success of BitTorrent has in its wake
attracted the attention of several researchers who examined its
performance in detail as highlighted in the previous section.
Several researchers subsequently suggested improvements to
existing BitTorrent mechanisms, or suggested new mecha-
nisms to replace existing BitTorrent mechanisms with the aim
of further improving its performance. In general, the suggested
improvements can be viewed as belonging to two categories:
(a) mechanisms that suggest improvements to BitTorrent’s
overlay topology formation and maintenance, and (b) mod-
ifications to the piece exchange mechanisms in BitTorrent. In
this section, we present an overview of the research literature
suggesting improvements to BitTorrent. We then categorize
these mechanisms as stated above.
A. Improvements to Overlay Topology
First, we review several proposals that aim to improve
the performance of BitTorrent by changing the way peers
establish their connections to other peers while joining, and
thereafter while maintaining the connections. These techniques
are aimed at addressing the topology mismatch problem be-
tween the overlay and the underlying physical network. The
techniques are based on using the location of the peers to
optimize the overlay topology. We classify the techniques into
two categories: (1) proximity aware techniques, and (2) ISP-
friendly techniques.
1) Proximity Awareness: BitTorrent builds its overlay net-
work by selecting neighbors randomly. The randomized se-
lection results in a significant divergence between the log-
ical overlay network and the underlying physical network
infrastructure [49]. This causes wastage of network resources
and potentially results in suboptimal downloading time. For
instance, two neighboring peers in the overlay may be in
different countries far away from each other. In addition,
selecting neighbors randomly also increases the probability
of having capacity-limited peers connected to high-capacity
peers. To address these problems, some researchers have
explored the method of using proximity in the topology
construction and maintenance in BitTorrent. The major aims
of exploiting proximity are achieving an efficient usage of
network resources, and reducing the individual downloading
time for the peers.
Qureshi [49] was among the first to suggest the use of
proximity in the BitTorrent overlay network. He pointed out
that the present overlay construction algorithms ignore the
problem of matching the overlay network and the underlying
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

152 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
TAB L E V
SUMMARY OF METHODS USED IN STUDYING BITTORRENT PERFORMANCE
Issues Measurement Simulation Analysis
Piece Exchange
TFT(Incentive
mechanism)
The TFT policy works well [18].
The replacement with a bit level
TFT solution is not appropriate
[22]. The TFT mechanism in-
creases the cooperative behavior
and discourages free-riding, but if
there are a large number of seeds, it
may not work effectively to prevent
free-riding [20].
The TFT policy fails to prevent
unfairness [29].
The incentive mechanism are
efficient [3]. Optimistic unchoking
may encourage free-riders [3].
BT’s mechanisms can guard
against free-riding successfully
[39]. TFT policy is not the major
cause for good performance; it
is the altruistic contribution by
high-capacity peers [42]
Piece Selection The rarest first policy works well
[18]. The rarest first policy pre-
vents the reappearance of rare
pieces and avoids the last pieces
problem [22].
Rarest first piece selection strategy
gives balanced performance [28].
The rarest first policy is critical to
ensure the new joined peers have
something to exchange with others
[29].
Rarest first strategy is still the best
waytoincreasefile availability
[45].
Overlay Topology
Topology Choking algorithm can cluster
peers with similar bandwidth [24].
There is no evidence of clustering
except the initial stage [25].
The overlay is not a random graph
[31] The overlay topology can be
affected by the choice of peer set
size and the percentage of outgoing
connections [32].
Neighbor
Selection
The random neighbor selection in
BT causes low transmission rate
and high operating cost [26]. BT is
locality-unaware, which increases
the ISPs’ cost [8].
The peer selection strategy may
lead to the mismatched peering and
make BT deviate more from the
optimum [33].
Impact of Fire-
wall/NAT
Peers behind a firewall have low
bandwidth usage [23]. Peers be-
hind a firewall often disconnect
early [23].
A large number of NATed peers de-
crease the robustness of the overlay
to attacks significantly [31]
Presence of non-firewalled peers
essential to good system perfor-
mance [45].
network. He suggested that peers that are close by in the real
world should be close by in the overlay network. Formally,
for any two nodes A and B, he specified their round-trip
communication latency as rtt[A,B]. Given three nodes A, B
and C, if rtt[A,B]<rtt
[A,C],thenP(A, B )>P(A, C ),
where P(A, B)is the probability that A and B are peers in the
overlay network. Several techniques including using network
coordinates coupled with a probabilistic flooding algorithm
were exploited in the topology construction and maintenance.
This was done in two phases: (1) estimating the node’s
network coordinates; (2) using a gossip-like protocol for near
neighbor discovery. The expected result was that an overlay
network that closely matches the underlying network topology
in terms of node proximity should yield shorter downloading
time and more efficient usage of network resources.
Zhang et al. [50] analyzed location proximity in BitTorrent
system not only based on topology proximity, but also for
cooperative proximity. They indicated that both neighbor
selection and piece selection phases are proximity-unaware.
This can cause problems such as topology mismatch, and long
average downloading time. They considered a proximity aware
topology construction algorithm. Their approach consists of
two steps. In the first step, they give each node a synthetic
network coordinate by using the algorithms in [51]. In the
second step, each node selects its neighbors by their Euclidean
distance. Their simulations are conducted using a modified
version of BTSim [34] extended to exploit proximity in over-
lay topology construction. They built a 3-dimensional space
representation of the network, where each node’s location was
specified as a 3-tuple representing the three coordinates, which
could be x-coordinate, y-coordinate and z-coordinate. When a
node joins into the network, the coordinates will be assigned
to it by the tracker. Then, the node will measure the Euclidean
distance between other nodes and itself, and proactively select
nodes with the shortest Euclidean distance as its neighbors.
Their results show that using location proximity in BitTorrent
can decrease the file downloading time about 11.3% for 1000
nodes torrent, and improve the network utilization, especially
for moderate size torrent.
Koo et al. [52] proposed a neighbor selection strategy
for hybrid P2P network like BitTorrent that favors picking
neighbors with the most mutually disjoint content. In [52],
they modeled peers as an undirected graph. In addition,
they presented a genetic-algorithm-based method for neigh-
bor selection. They use ns-2 [53] package to simulate the
system. In their simulation model, there are three classes of
peers with different bandwidths. They compared the proposed
neighbor selection strategy with the existing strategy using two
metrics: system throughput and average downloading time.
Their results show that the new neighbor selection algorithm
can improve system performance by increasing the content
availability for peers from their neighbors.
The proposed methods of neighbor selection by proximity
[49], [50] are still oblivious to ISP boundaries. Peers selected
with proximity in mind may belong to different ISPs, thus still
not addressing the cross-ISP traffic problem.
2) ISP-Friendliness: The wide use of BitTorrent has put
ISPs in a dilemma [54]! On the one hand, BitTorrent, with
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 153
a large number of users, constitutes an important source of
revenue for ISPs. On the other hand, BitTorrent generates a
large amount of cross-ISP traffic and makes the ISPs’ cost
increase significantly since BitTorrent’s implementations ig-
nore the underlying Internet topology and ISP costs. Lei et al.
[26] mentioned that BitTorrent’s fast downloading experience
comes at a large cost to the network in terms of resources. This
may consequently affect other Internet services supported by
the ISPs.
To address the cross-ISP problem, Bindal et al. [9] designed
a new algorithm based on biased neighbor selection, in which
a peer chooses its neighbors mostly from peers within the
same ISP, and only selects a few from the outside its ISP
domain. This is because peers in the same ISP are highly
connected. However, some connectivity to peers outside the
ISP boundaries needs to be maintained in order to obtain new
blocks. They indicated that the biased neighbor selection can
be implemented either by changing the tracker and client or by
employing P2P traffic shaping devices. When there is a request
from a peer for neighbor list, the tracker will choose kexternal
peers and 35 −kinternal peers to feed back to the requesting
peer. If there are less than 35 −kinternal peers, the tracker
will ask the requesting peer to contact it again after certain
duration. The tracker can know the peers’ ISP location either
by using the AS mapping or IP address to identify the ISP,
or add a new header which contains the locality tag to HTTP
proxy. Biased neighbor selection can also be implemented by
P2P traffic shaping devices, which are situated alongside the
edge routers of the ISPs and use deep packet inspection to
identify P2P traffic. To understand the process of neighbor
selection, the authors designed a discrete-event simulator and
evaluated it over a network configuration with 14 ISPs. They
considered the flash crowd phase. The main metrics evaluated
include the download time, and ISP traffic redundancy which
is the average number of times the blocks of the downloading
file travels into the ISP until all peers inside the ISP finish their
download. The simulation result shows that biased neighbor
selection can reduce the cross-ISP traffic and maintain the
nearly optimal performance of BitTorrent.
Yamazaki et al. [55] proposed a series of strategies to reduce
ISP costs called Cost-Aware BitTorrent (CAT). In CAT, the
peers first acquire ISP cost information. Thereafter, path costs
between any two peers can be computed using the ISP costs.
The tracker selects peers in the list of peers it returns based
on the shortest cost to the requesting peer. The inter-peer
cost is used by a peer in the selection of candidate peers
from which to download and upload pieces. The strategy is to
minimize the cost incurred in the transactions, consequently
minimizing the ISP costs. Through simulations for different
scenarios they show that CAT can reduce the ISP cost and
improve the performance.
B. Improvements to Piece Exchange Mechanism
As already mentioned in Sec. II, the tit-for-tat (TFT) mech-
anism ensures that peers contribute as much as they download.
Similarly, optimistic unchoking (OU) enables new peers to get
started, and enables peers to discover better matched peers for
cooperation. In this section, we review several suggestions by
authors to improve the piece exchange mechanisms in BitTor-
rent to further improve download performance, punish free-
riders and address unfairness. The first set of mechanisms aims
to promote collaboration among peers using the underutilized
download bandwidth of peers to help others. The second set
of mechanisms address the issue of free-riding through several
approach to identify and punish free-riders. The third set of
mechanisms address the issue of unfairness in BitTorrent by
suggesting improvements to the unchoke mechanisms.
1) Collaboration among Peers: BitTorrent’s piece ex-
change mechanism is designed to enforce fairness among
peers through the TFT mechanism. Garbacki et al. [56]
pointed out three limitations of TFT: (1) the newcomers are
bootstrapped at the bandwidth cost of the existing peers; (2)
no incentives for seeding; (3) peers with asymmetric Internet
connections cannot fully use their downloading links since
they are forced to download at the speed of their uploading
link.
Garbacki et al. [57] proposed a protocol named 2Fast which
extended the bartering model of BitTorrent. In 2Fast, a peer
with idle bandwidth can join a download in progress as a
helper to assist a peer (hereafter named the collector)andfetch
missing fragments of the file being downloaded and thereafter
upload them to the collector. A collector recruits peers that are
willing to act as its helpers with the promise that the bandwidth
contributed by the helper will be returned in the future. The
authors used social incentives to enforce the delivery on the
promise to return bandwidth in the 2Fast system. In [58], the
same authors exploited social phenomena like friendships to
enable collaboration. Here a collector finds helpers just like
people collaborate with friends in communities. Through rig-
orous analysis, the authors show that adding a new helper will
not help if the collector’s downloading bandwidth is already
saturated. They computed the minimum number of helpers
that fills the collector’s bandwidth. Finally, they conducted
experiments to compare the BitTorrent protocol enhanced with
the 2Fast protocol against the standard BitTorrent protocols
using several metrics such as the downloading speedup, which
is the ratio between the downloading time of a peer acting on
its own and the downloading time of a collector supported by
its helpers. The results show that the use of 2Fast protocol
improves the download speed by up to a factor of 3.5 in
comparison to the standard BitTorrent without helpers. Fig. 3
depicts the collaboration between a collector and its helpers
in the 2Fast protocol.
Subsequently, Garbacki et al. [56] extended the 2Fast pro-
tocol and proposed a novel mechanism in which incentives
are built around bandwidth rather than content. Bandwidth
is unrelated to the interests of peers, so it is more suitable
to be a trading unit. In the bandwidth-exchange incentive
mechanism, helpers (peers) can use their idle bandwidth to
help others with their download, regardless of the content.
The new protocol is named amortized tit-for-tat (ATFT). In
this protocol peers choose the bandwidth borrowers with the
highest chance of returning the borrowed bandwidth. ATFT
employs the exploration and selection operations to select
peers with the highest contributions as the borrowers. Explo-
ration extends the borrowers set by opening opportunities for
bandwidth exchange, and selection reduces the borrowers set
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

154 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
Collaborative Download Non-collaborative Download
Helper
Collector Dowload Completed
Downloading
Fig. 3. Collector and helpers in 2Fast system
by removing the least promising peers. Fig. 4 illustrates the
ATFT protocol from the view of the collector. The authors
pointed out four benefits of ATFT. Firstly, newcomers pay the
cost of bootstrapping by themselves. Secondly, ATFT provides
an effective incentive for content seeding. Thirdly, peers can
fill their downlinks completely with the helpers’ bandwidth.
Finally, it can be used to provide a certain level of anonymity.
They conducted several simulation experiments using ATFT.
In their simulations, a peer maintains a single value for
keeping track of the bandwidth obtained from contributors to
account for bandwidth contributions. The results show that
ATFT improves the average download bandwidth of a peer by
a factor of 2 to 6.
The above protocols do not explicitly specify how many
pieces the helpers should download to make the system work
well. On one hand, if helpers download too many pieces, it
would incur excessive transmission cost; on the other hand,
if the helper downloads too few pieces, it may not be very
useful. Wang et al. [59] showed that helpers can work as
effectively as seeders if they download only a tiny fraction
of the file. Helpers in [59] are resource-rich nodes with spare
uploading bandwidth that can be used to increase the total
system uploading bandwidth and hence easing the bottleneck.
They proposed an efficient strategy for utilizing the spare
uploading capacity, where a helper joins a swarm just like
a regular peer, and only downloads a small number, kpieces,
of the file. Once a helper downloads kpieces, it will stop
downloading, and at this time becomes a microseed.The
microseed contacts the tracker and gets a list of peers who
need help. The microseed implements choking and unchoking
algorithms. Once the microseed knows that a neighbor has
already downloaded all the pieces that it possesses, it will
disconnect from the neighbor. Helpers download only a small
number of pieces because they are not interested in the file.
Likewise, downloading a large part of the file is a waste of
helpers’ bandwidth. The authors used a fluid model based on
the model presented in [3] to analyze the system performance.
They made several assumptions: the peer arrival process is
Poisson; peers are homogeneous with the same uploading
and downloading bandwidth; peers download at the same rate
Non-cont ributor peer
Requesting (busy)
contributor
Contrib utors set
Collaboration
Tit-for-tat
Idle contributor acting as
helper
Collect or
Contribution from helper
Fig. 4. The Collector and helpers in ATFT protocol
when they are downloading. They also implemented a simu-
lator for the original BitTorrent and BitTorrent with helpers,
using which they compared the average downloading time.
They presented several conclusions: (1) helpers are almost as
effective as seeds though they only download a tiny fraction
of the file; (2) helpers that download too many pieces will hurt
the system performance; (3) average downloading time of a
BitTorrent system with helpers is much shorter than that of
the BitTorrent system without helpers; (4) it is more effective
for BitTorrent to have a lot of helpers than to have a few extra
seeds.
2) Addressing Free-Riding: A free-rider is a node that
downloads pieces from other peers but does not upload any
pieces to others. Conventional P2P systems lack incentive
mechanisms and are vulnerable to free-riding. Incentive mech-
anisms encourage cooperation among peers by rewarding
contributors and punishing free-riders. Although BitTorrent
has incentive mechanisms such as TFT, many researchers
pointed out that they are not effective in preventing free-riding.
Jun et al. [60] evaluated the original incentive mecha-
nism of BitTorrent using measurements on a P2P network
set up on PlanetLab [61]. They view the peer behavior in
BitTorrent as akin to the iterated prisoner’s dilemma. They
found that free-riders are not effectively punished, and peers
that contribute to others are not rewarded appropriately. To
address this problem, they proposed a new mechanism which
is more robust against free-riders. In their new mechanism,
they defined peers’ upload amount uand download amount
dfor each link. Then, they define a concept named deficit,
which is u−d. A peer ensures that the deficit is always upper-
bounded by the formula: u−d≤f.c, where the constant c
is the size of a fragment, and f(≥1) is called a nice factor.
The factor determines the amount that a peer is willing to risk
for a chance to establish cooperation. Peers upload evenly
to all links as much as they can under this condition. The
authors evaluated the new mechanism experimentally through
measurements using PlanetLab. The experiments consisted of
one tracker, one seed and 170 leechers running BitTorrent
4.0.0. All of the leechers begin to download at the same time.
During the experiment, the authors evaluated two metrics:
the downloading time and the uploading amount. The results
show that the original incentive mechanism in BitTorrent is
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 155
suboptimal and the proposed mechanism is more robust at
preventing free-riding.
Sirivianos et al. [62] presented a new free-riding technique
named the large view exploit. In this approach they modified
the BitTorrent client to: (a) not upload content to others, (b)
periodically contact the tracker behaving as a new peer and
obtain a new list of peers, and (c) establish connections with
peers in the new list, thus exploiting their optimistic unchoke
contributions. The authors performed several experiments both
with PlanetLab resident torrents and public torrents to study
the effectiveness of the large view exploit. Their results show
that the modified peer can easily achieve better performance
than the traditional peers. The authors then suggest a modi-
fication to the BitTorrent tracker and clients to address this
problem. The aim is to give the clients and the tracker a
consistent view of the whole swarm. The approach is to use
a peer’s IP address together with a pseudo-random number
sequence to enable other peers to attempt to identify free-
riders. The modification requires the tracker to perform a
larger amount of housekeeping, and may impact the scalability
of the approach. The authors are continuing to investigate this
approach further to improve scalability.
Chow et al. [63] presented a novel approach from the
perspective of using the seed capacity appropriately with the
goal of reducing the free-riders. Several of the measurements
and simulation studies of BitTorrent reveal that the leechers’
downloading rate is slow at the beginning when they have
few chunks to exchange with others. Similarly, it can be
slow at the end due to the inability of the peer in finding
neighbors possessing the few missing chunks that it requires
to complete the download. So the authors proposed a simple
method to prioritize the use of seeding capacity to certain
portions of the file downloading process. They used two
ways to choose neighbors to unchoke: (1) Sort-based: a seed
sorts its neighbors according to the number of chunks each
possesses. Then it unchokes Nof them based on the sorting
order; (2) Threshold-based: a seed unchokes Nneighbors
with [0..K/2]%or[(100 −K/2)..100]% of the chunks. Their
experiments use BTSim [9] with the following modifications:
(1) nodes could be seeds with different seeding approach; (2)
nodes act as free-riders; (3) continuous node arrivals. Their
experimental results show that the new seed capacity prioriti-
zation approach not only discourages free-riding behavior, but
also improves the performance of contributing leechers.
3) Improving Fairness: Bharambe et al. [29] found that
BitTorrent’s optimistic unchoke mechanism can result in sys-
tematic unfairness where a high-capacity node might end
up uploading a larger amount of data than it downloads.
They proposed two strategies to reduce this unfairness. The
first one named Quick bandwidth estimation (QBE) tries to
circumvent the need for optimistic unchokes. This is feasible
if the nodes can quickly estimate the uploading bandwidths
for all its neighboring peers. QBE can estimate a neighbor’s
upload bandwidth based on the history of past interactions
with it. Alternately, the packet-pair principle can be used to
estimate the bandwidth. The second proposed algorithm is
pairwise block-level TFT, which focuses on fairness of the
number of blocks transferred instead of uploading rates. Node
Aallows its neighbor Bto download a block from it if
and only if Uab ≤Dab +Δ ,whereUab and Dab are the
uploaded and the downloaded blocks from node Ato/from
node Brespectively, and Δis the unfairness threshold for
this connection. This ensures that the maximum number of
extra blocks served by a node is bounded by dΔwhere dis
the size of its neighborhood. The authors used their simulator
to evaluate the performance of the new techniques based
on three metrics: mean uploading utilization, unfairness, and
mean downloading time. The results showed that the proposed
methods, QBE and pairwise block-level TFT can effectively
address the unfairness towards high-capacity nodes, and also
improve the upload link utilization. We should note that QBE
is somewhat idealistic since the reliable bandwidth estimation
is far from a trivial exercise [64]. The block-level TFT may
result in a reduction of uploading utilization because peers
can potentially cease to upload when the block-level TFT
constraint is not satisfied.
Thommes et al. [64] simulated a simplified version of
BitTorrent protocol using Matlab in order to understand the
fairness properties. Based on their evaluation, they proposed
three methods to improve BitTorrent fairness. The three new
mechanisms are as follows: The first method is conditional
optimistic unchoke. In this method a peer performs an op-
timistic unchoke only when it’s IFR is larger than 1. The
second method is multiple connection chokes, which allows
peers to choke/unchoke not just one connection, but multiple
connections in each round. A peer computes the Connection
Fairness, which is the ratio of the peer’s uploading rate to
a specific connected peer to the downloading rate from that
peer, for each of the five peers to which it is connected. They
evaluate two parameters: (a) Threshold Ratio, which is the
largest value the Connection Fairness can assume before the
corresponding upload is choked; (b) Maximum Chokes (MC),
which is the largest number of uploads a peer can choke
in each round. If the number of connection fairness values
exceeding the Threshold Ratio is larger than the Maximum
Chokes, the peer will choke the connections which are unfair.
Otherwise, it will choke only MC of the peers. The third mech-
anism is variable number of outgoing connections, in which
the number of connections a peer maintains is variable, instead
of a fixed number. The number of outgoing connections varies
depending on the peer’s upload capacity. They also allow the
peer’s upload capacity dedicated to each connection to be
variable. Each peer evaluates its set of outgoing and incoming
connections every 10 seconds. In each iteration, the peer
constructs a list Lnd of peers to which it is currently uploading,
but from which the peer is not receiving any data. The peer
will choke the peers in list Lnd. Then, the peer constructs
another list Lnu of peers from which it is downloading, but to
which it is not uploading. When |Lnu |is larger than |Lnd|,it
starts uploading to a random set of |Lnd |peers in Lnu .When
|Lnu|is smaller than |Lnd|, it begins to upload to all peers
in Lnu, and randomly choose |Lnd |−|Lnu|peers to unchoke
optimistically.
They defined the Instantaneous Fairness Ratio (IFR) for
each peer as the ratio of data uploaded to data downloaded
during the previous 10 seconds. If a peer’s IFR is less than 1,
it means the peer is downloading an excessive amount. Oth-
erwise, it tells us that the peer is downloading an insufficient
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

156 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
amount. They evaluated the three modifications by simulation.
The results show that the three modifications provide some
level of improvement in BitTorrent’s fairness.
C. Summary
Table VI summarizes the improvements to BitTorrent’s
mechanisms that are presented earlier in this section. As can
be seen from the table, each suggested improvement addresses
one of the issues with BitTorrent mechanisms. No studies of
how these improvements interact with each other if included
together in BitTorrent are available. It would be interesting to
explore the combined effects of the different improvements
suggested to see if they indeed work collaboratively to further
improve performance, or impede each other.
V. C ONCLUSION
This paper presented a survey of BitTorrent performance.
First we reviewed the performance studies of the original
BitTorrent’s protocols including measurement, simulation, and
analytical modeling. We then summarized the findings of
these studies. Next, we summarized some of the suggested
improvements to BitTorrent’s mechanisms in order to further
improve its performance.
We note that the innovative mechanisms introduced by Bit-
Torrent has influenced not only P2P file sharing applications,
but also served as the foundation for several other applications
including P2P streaming and P2P Voice over IP. For example,
BitTorrent based streaming applications are described in [65],
[66]. Thus we see the study of BitTorrent and its performance
as a very useful step not only to understand its performance,
but also as a means of learning from the experience to enable
better design of similar protocols for other applications.
ACKNOWLEDGMENT
The work described in this paper has been supported by
HK RGC under RGC-Competitive Earmarked Research Grant
HKUST 617907. The authors would like to thank all the
anonymous reviewers for their invaluable suggestions.
REFERENCES
[1] “Bittorrent,” http://www.bittorrent.com/.
[2] B. Cohen, “Incentives build robustness in BitTorrent,” in First Workshop
on Economics of Peer-to-peer Systems, Berkeley, USA, June 2003.
[3] D. Qiu and R. Srikant, “Modeling and performance analysis of
BitTorrent-like peer-to-peer networks,” in SIGCOMM ’04: Proc. 2004
conference on Applications, technologies, architectures, and protocols
for computer communications. New York, NY, USA: ACM, 2004, pp.
367–378.
[4] “Gnutella,” http://www.gnutelliums.com/.
[5] “Kazaa,” http://www.kazaa.com/.
[6] J. A. Pouwelse, P. Garbacki, D. H. J. Epema, and H. J. Sips, “The
BitTorrent p2p file-sharing system: Measurements and analysis,” in
IPTPS’05, 2005, pp. 205–216.
[7] L. Ellis, “BitTorrent’s swarms have a deadly bite on broadband nets,”
Multichannel news, 2006.
[8] T. Karagiannis, P. Rodriguez, and K. Papagiannaki, “Should internet ser-
vice providers fear peer-assisted content distribution?” in IMC ’05: Proc.
5th ACM SIGCOMM conference on Internet Measurement.Berkeley,
CA, USA: USENIX Association, 2005, pp. 6–6.
[9] R. Bindal, P. Cao, W. Chan, J. Medved, G. Suwala, T. Bates, and
A. Zhang, “Improving traffic locality in BitTorrent via biased neighbor
selection,” in ICDCS ’06: Proc. 26th IEEE International Conference
on Distributed Computing Systems. Washington, DC, USA: IEEE
Computer Society, 2006, p. 66.
[10] C. Gkantsidis and P. Rodriguez, “Network coding for large scale content
distribution,” in IEEE INFOCOM. Miami, USA: IEEE Computer
Society, March 2005, pp. 2235–2245.
[11] D. Levin, R. Sherwood, and B. Bhattacharjee, “Fair file swarming
with FOX,” in In Fifth International Workshop on Peer-to-peer Systems
(IPTPS 2006), 2006.
[12] R. Sherwood, R. Braud, and B. Bhattacharjee, “Slurpie: a cooperative
bulk data transfer protocol,” INFOCOM 2004. Twenty-third AnnualJoint
Conference of the IEEE Computer and Communications Societies,
vol. 2, pp. 941–951, March 2004.
[13] E. K. Lua, J. Crowcroft, M. Pias, R. Sharma, and S. Lim, “A survey and
comparison of peer-to-peer overlay network schemes,” IEEE Commun.
Surveys Tutorials, vol. 7, no. 2, pp. 72–93, Quarter 2005.
[14] S. Androutsellis-Theotokis and D. Spinellis, “A survey of peer-to-peer
content distribution technologies,” ACM Comput. Surv., vol. 36, no. 4,
pp. 335–371, 2004.
[15] M. Hosseini, D. Ahmed, S. Shirmohammadi, and N. Georganas, “A
survey of application-layer multicast protocols,” IEEE Commun. Surveys
Tutor ia ls , vol. 9, no. 3, pp. 58–74, Quarter 2007.
[16] D. Chopra, H. Schulzrinne, E. Marocco, and E. Ivov, “Peer-to-peer
overlays for real-time communication: Security issues and solutions,”
IEEE Commun. Surveys Tutorials, vol. 11, no. 1, pp. 4–12, 2009.
[17] R. Ranjan, A. Harwood, and R. Buyya, “Peer-to-peer-based resource
discovery in global grids: A tutorial,” IEEE Commun. Surveys Tutorials,
vol. 10, no. 2, pp. 6–33, 2008.
[18] M. Izal, G. Urvoy-keller, E. W. Biersack, P. A. Felber, A. A. Hamra,
and L. Garces-Erice, “Dissecting BitTorrent: Five months in a torrents
lifetime,” in Passive and Active Measurements (PAM), LNCS, vol. 3014,
April 2004, pp. 1–11.
[19] A. Bellissimo, B. N. Levine, and P. Shenoy, “Exploring the use of
BitTorrent as the basis for a large trace repository,” Technical report,
University of Massachusetts Amherst, Dept. of Computer Science, Tech.
Rep., June 2004.
[20] N. Andrade, M. Mowbray, A. Lima, G. Wagner, and M. Ripeanu,
“Influences on cooperation in BitTorrent communities,” in P2PECON
’05: Proceedings of the 2005 ACM SIGCOMM workshop on Economics
of peer-to-peer systems. New York, NY, USA: ACM, 2005, pp. 111–
115.
[21] A. Rasti and R. Rejaie, “Understanding peer-level performance in
BitTorrent: A measurement study,” Computer Communications and
Networks, 2007. ICCCN 2007. Proceedings of 16th International Con-
ference on, pp. 109–114, Aug. 2007.
[22] A. Legout, G. Urvoy-Keller, and P. Michiardi, “Rarest first and choke
algorithms are enough,” in IMC ’06: Proc. 6th ACM SIGCOMM
conference on Internet measurement. New York, NY, USA: ACM,
2006, pp. 203–216.
[23] K. andr ´e Skevik, V. Goebel, and T. Plagemann, “Analysis of BitTorrent
and its use for the design of a p2p based streaming protocol for a hybrid
cdn,” Delft University of Technology, Tech. Rep., 2004.
[24] A. Legout, N. Liogkas, E. Kohler, and L. Zhang, “Clustering and sharing
incentives in BitTorrent systems,” in SIGMETRICS ’07: Proceedings of
the 2007 ACM SIGMETRICS international conference on Measurement
and modeling of computer systems. New York, NY, USA: ACM, 2007,
pp. 301–312.
[25] C. Dale, J. Liu, J. Peters, and B. Li, “Evolution and enhancement of
BitTorrent network topologies,” Quality of Service, 2008. IWQoS 2008.
16th International Workshop on, pp. 1–10, June 2008.
[26] Y. Lei, L. Yang, Q. Jiang, and C. Wu, “Experimental views on neighbor
selection in BitTorrent,” Network and Parallel Computing Workshops,
2007. NPC Workshops. IFIP International Conference on, pp. 813–818,
Sept. 2007.
[27] “Bittornado website,” http://www.bittorrent.com/, 2007.
[28] P. A. Felber and E. W. Biersack, “Self-scaling networks for content dis-
tribution,” in Int. Workshop on Self-* Properties in Complex Information
Systems, Bertinoro, Italy, May-June 2004.
[29] A. R. Bharambe, C. Herley, and V. N. Padmanabhan, “Analyzing and
improving a BitTorrent networks performance mechanisms,” INFOCOM
2006. 25th IEEE International Conference on Computer Communica-
tions. Proceedings, pp. 1–12, April 2006.
[30] E. Adar, “Drawing crowds and bit welfare,” SIGecom Exch.,vol.5,
no. 4, pp. 31–40, 2005.
[31] A. Al Hamra, A. Legout, and C. Barakat, “Understanding the properties
of the BitTorrent overlay,” INRIA, Sophia Antipolis, Tech. Rep., July
2007.
[32] G. Urvoy-Keller and P. Michiardi, “Impact of inner parameters and
overlay structure on the performance of BitTorrent,” INFOCOM 2006.
25th IEEE International Conference on Computer Communications.
Proceedings, pp. 1–6, April 2006.
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 157
TAB L E V I
SUMMARY OF IMPROVED MECHANISMS SUGGESTED FOR BITTORRENT
Mechanisms Suggested Improvements Features of Improvement Result Ref.
Overlay Topology
Random
Neighbor
Selection
Neighbor selection by
proximity
Synthetic network coordinates, Eu-
clidean distance
Decrease the file downloading time
and increase network utilization
[49],
[50]
Neighbor selection by con-
tent
Select neighbor by the most mutually
disjoint content
Maximize content availability from
neighbors
[52]
ISP-
unawareness
Biased neighbor selection Choose most neighbors inside the ISP Reduce the cross-ISP traffic[9]
Cost-aware Selection Select neighbors according to the cost
between peers
Reduce the ISP cost [55]
Piece Exchange
No Collabo-
rative Mech-
anism
Collaborative Download Helpers: peers with idle bandwidth can
help other peers downloading
Improve download speed; better re-
source utilization
[57],
[58],
[59]
ATFT Incentives are built on bandwidth in-
stead of content
Several benefits including for new
peers, seeds; better resource uti-
lization
[56]
Helpers Helpers download a small fraction of
file and act as microseeds
Improved download performance
for peers
[59]
Addressing
Free-Riding
New incentive mechanism Peers upload evenly to all links as
much as they can under the condition
u−d<=f.c.
The new proposed mechanism is
more robust to prevent free-riding
[60]
Addressing the problems
of large view exploit
Using peer’s IP address to track free-
riding behavior
Better ability to identify free-riders [62]
Seed capacity utilization Use the seeding capacity to certain por-
tions of a file downloading process
Discourage free-riders’ behavior,
Improve the leechers’ Performance
[63]
Improving
Fairness
QBE; Pairwise block-level
TFT
QBE: uploading bandwidths can be es-
timated quickly; Pairwise block-level
TFT: the maximum number of extra
blocks served by a node is bounded
Address the problems of low up-
link utilization and unfairness
[29]
Conditional optimistic un-
choke; Multiple connec-
tion chokes; Variable num-
ber of outgoing connec-
tions
Conditional optimistic unchoke: only
when a peer’s IFR is larger than 1,
it performs an optimistic unchoke al-
gorithm; Multiple connection chokes:
peers can choke/unchoke multiple con-
nections in each round; Variable number
of outgoing connections: the number of
connections a peer maintains is variable
Improve fairness [64]
[33] G. Wu and T. cker Chiueh, “How efficient is BitTorrent?” in Multimedia
Computing and Networking 2006, S. Chandra and C. Griwodz, Eds.,
vol. 6071, no. 1. SPIE, 2006, p. 60710O. [Online]. Available:
http://link.aip.org/link/?PSI/6071/60710O/1
[34] A. Bharambe, C. Herley, and V. Padmanabhan, “Mi-
crosoft research simulator for the BitTorrent protocol,”
http://research.microsoft.com/projects/btsim/, 2006.
[35] X. Yang and G. de Veciana, “Performance of peer-to-
peer networks: Service capacity and role of resource sharing
policies,” Performance Evaluation, vol. 63, no. 3, pp. 175
– 194, 2006, p2P Computing Systems. [Online]. Avail-
able: http://www.sciencedirect.com/science/article/B6V13-4FJGW2F-
1/2/be7c36ff2970dd2cea798f2de6200a40
[36] L. Guo, S. Chen, Z. Xiao, E. Tan, X. Ding, and X. Zhang, “A
performance study of BitTorrent-like peer-to-peer systems,” IEEE J. Sel.
Areas Commun., vol. 25, no. 1, pp. 155–169, Jan. 2007.
[37] D. Arthur and R. Panigraphy, “Analyzing the efficiency of BitTorrent
and related peer-to-peer networks,” in Proc. Seventeenth annual ACM-
SIAM symposium on Discrete algorithms, January 2005, pp. 961–969.
[38] R. Kumar and K. Ross, “Peer-assisted file distribution: The minimum
distribution time,” Hot Topics in Web Systems and Technologies, 2006.
HOTWEB ’06. 1st IEEE Workshop on, pp. 1–11, Nov. 2006.
[39] J. Yu, M. Li, F. Hong, and G. Xue, “Free-riding analysis of BitTorrent-
like peer-to-peer networks,” Services Computing, 2006. APSCC ’06.
IEEE Asia-Pacific Conference on, pp. 534–538, Dec. 2006.
[40] Y. Tian, D. Wu, and K. W. Ng, “Modeling, analysis and improvement
for BitTorrent-like file sharing networks,” INFOCOM 2006. 25th IEEE
International Conference on Computer Communications. Proceedings,
pp. 1–11, April 2006.
[41] P. Saraswat and P. Batra, “An empirical performance evaluation and
modelling of BitTorrent peer-to-peer file sharing system using queuing
network models,” Research project, Advanced Learning and Research
Institute, Switzerland, 2007.
[42] M. Piatek, T. Isdal, T. Anderson, A. Krishnamurthy, and A. Venkatara-
mani, “Do incentives build robustness in BitTorrent?” in NSDI’07,
Cambridge, MA, April 2007, pp. 1–14.
[43] F. Lo Piccolo and G. Neglia, “The effect of heterogeneous link capacities
in BitTorrent-like file sharing systems,” Peer-to-Peer Systems, 2004.
International Workshop on Hot Topics in, pp. 40–47, Oct. 2004.
[44] F. Cl´evenot-Perronnin, P. Nain, and K. W. Ross, “Multiclass
p2p networks: Static resource allocation for service differentiation
and bandwidth diversity,” Performance Evaluation,vol.62,
no. 1-4, pp. 32 – 49, 2005, performance 2005. [On-
line]. Available: http://www.sciencedirect.com/science/article/B6V13-
4GV9S5V-5/2/182bf23d02fcbeff01e45a97ed891a9a
[45] B. Fan, D.-M. Chiu, and J. Lui, “Stochastic analysis and file availability
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.

158 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010
enhancement for BT-like file sharing systems,” Quality of Service, 2006.
IWQoS 2006. 14th IEEE International Workshop on, pp. 30–39, June
2006.
[46] J. Mundinger, R. Weber, and G. Weiss, “Optimal scheduling of peer-
to-peer file dissemination,” Journal of Scheduling, vol. 11, no. 2, pp.
105–120, April 2008.
[47] W.-C. Liao, F. Papadopoulos, and K. Psounis, “Performance analysis
of BitTorrent-like systems with heterogeneous users,” Performance
Evaluation, vol. 64, no. 9-12, pp. 876 – 891, 2007. [On-
line]. Available: http://www.sciencedirect.com/science/article/B6V13-
4P192CX-2/2/d13c7120e07234e69160c85b5b861f8d
[48] M. Bertoli, G. Casale, and G. Serazzi, “An overview of the jmt queuing
network simulator,” Politecnico di Milano, Tech. Rep., 2007.
[49] A. Qureshi, “Exploring proximity based peer selection in a BitTorrent-
like protocol,” MIT 6.824 student project, 2004.
[50] L. Zhang, J. K. Muppala, and W. Tu, “Exploiting proximity in coop-
erative download of large files in peer-to-peer networks,” International
Conference on Internet and Web Applications and Services, (ICIW’07),
vol. 0, p. 1, 2007.
[51] M. Pias, J. Crowcroft, S. Wilbur, S. Bhatti, and T. Harris, “Lighthouses
for scalable distributed location,” in Second International Workshop on
Peer-to-Peer Systems (IPTPS ’03), Feb 2003, pp. 278–291.
[52] S. G. Koo, K. Kannan, and C. G. Lee, “On neighbor-selection
strategy in hybrid peer-to-peer networks,” Future Generation
Computer Systems, vol. 22, no. 7, pp. 732 – 741, 2006. [On-
line]. Available: http://www.sciencedirect.com/science/article/B6V06-
4JRKCPC-2/2/e800f30606343b0a99ebf78c36033a68
[53] “Ns-2, the network simulator,” http://www.isi.edu/nsnam/ns/, 2003.
[54] V. Aggarwal, A. Feldmann, C. Scheideler, and M. Faloutsos, “Can isps
and p2p users cooperate for improved performance,” ACM SIGCOMM
Computer Communication Review, vol. 37, pp. 29–40, 2007.
[55] S. Yamazaki, H. Tode, and K. Murakami, “CAT: A cost-aware BitTor-
rent,” in 32nd IEEE Conference on Local Computer Networks (LCN
2007), Oct 2007, pp. 226–227.
[56] P. Garbacki, D. Epema, and M. van Steen, “An amortized tit-for-tat
protocol for exchanging bandwidth instead of content in p2p networks,”
Self-Adaptive and Self-Organizing Systems, 2007. SASO ’07. First
International Conference on, pp. 119–128, July 2007.
[57] P. Garbacki, A. Iosup, D. Epema, and M. van Steen, “2fast: Collaborative
downloads in p2p networks,” in P2P ’06: Proc. Sixth IEEE International
Conference on Peer-to-Peer Computing. Washington, DC, USA: IEEE
Computer Society, 2006, pp. 23–30.
[58] J. Pouwelse, P. Garbacki, J. Wang, A. Bakker, J. Yang, A. Iosup,
D. Epema, M. Reinders, M. van Steen, and H. Sips, “Tribler: A social-
based peer-to-peer system,” Concurrency and Computation: Practice &
Experience, vol. 20, no. 2, pp. 127–138, Feb 2008.
[59] J. Wang, C. Yeo, V. Prabhakaran, and K. Ramchandran, “On the role
of helpers in peer-to-peer file download systems: Design, analysis and
simulation,” in IPTPS’07, 2007.
[60] S. Jun and M. Ahamad, “Incentives in BitTorrent induce free riding,” in
P2PECON ’05: Proc. 2005 ACM SIGCOMM workshop on Economics of
peer-to-peer systems. New York, NY, USA: ACM, 2005, pp. 116–121.
[61] “Planetlab,” http://www.planet-lab.org/.
[62] M. Sirivianos, J. Park, R. Chen, and X. Yang, “Freeriding in BitTorrent
networks with the large view exploit,” in IPTPS’07, 2007.
[63] A. L. Chow, L. Golubchik, and V. Misra, “Improving BitTorrent: A
simple approach,” in Proceedings of IPTPS, 2008.
[64] R. Thommes and M. Coates, “BitTorrent fairness: analysis and improve-
ments,” in Workshop Internet, Telecom. and Signal Proc. (WITSP’05),
Dec 2005.
[65] S. Tewari and L. Kleinrock, “Analytical model for BitTorrent-based
live video streaming,” Consumer Communications and Networking
Conference, 2007. CCNC 2007. 4th IEEE, pp. 976–980, Jan. 2007.
[66] A. Vlavianos, M. Iliofotou, and M. Faloutsos, “BiToS: Enhancing
BitTorrent for supporting streaming applications,” INFOCOM 2006.
25th IEEE International Conference on Computer Communications.
Proceedings, pp. 1–6, April 2006.
Raymond Lei. Xia is a M. Phil. student in The Hong Kong University of
Science and Technology, Clear Water Bay, Kowloon, Hong Kong.
Jogesh K. Muppala (M’85-SM’02) received the Ph. D. degree in Electrical
Engineering from Duke University, Durham, NC in 1991. He is currently an
associate professor in the Department of Computer Science and Engineering,
The Hong Kong University of Science and Technology, Clear Water Bay,
Kowloon, Hong Kong. He was previously a Member of the Technical Staff
at Software Productivity Consortium (Herndon, Virginia, USA) from 1991 to
1992, where he was involved in the development of modeling techniques
for systems and software. While at Duke University, he participated in
the development of two modeling tools, the Stochastic Petri Net Package
(SPNP) and the symbolic Hierarchical Automated Reliability and Performance
Evaluator (SHARPE), both of which are being used in several universities
and industry in the USA. He was the program co-chair for the 1999 Pacific
Rim International Symposium on Dependable Computing held in Hong Kong
in December 1999. He also co-founded and organized The Asia-Pacific
Workshop on Embedded System Education and Research. He has also served
on program committees of many international conferences. He received
the Excellence in Teaching Innovation Award 2007. He was also awarded
the Teaching Excellence Appreciation Award by the Dean of Engineering,
HKUST. Dr. Muppala is a Senior Member of IEEE, IEEE Computer Society
and IEEE Communications Society, and a participating representative from
HKUST with EDUCAUSE.
Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 05,2010 at 08:59:25 UTC from IEEE Xplore. Restrictions apply.