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 a downloading function that enables the peer to download
the most popular approach for sharing large files using the the content once it is located. Traditional P2P file sharing
peer-to-peer paradigm. BitTorrent introduced several innovative systems such as Gnutella [4] and KaZaa [5] focus on quickly
mechanisms such as tit-for-tat (TFT) and rarest first to enable
efficient distribution of files among the participating peers. locating the peers that possess a given file. Once located,
Several studies examining the performance of BitTorrent and the file is directly downloaded from one of the peers having
its mechanisms have been published in the literature. In this the file. BitTorrent, on the other hand relies on cooperation
paper, we present a survey of performance studies of BitTorrent among the peers for file distribution using swarming [2].
from 2003 to 2008. We categorize these studies based on the BitTorrent differs from the conventional P2P file sharing
techniques used, the mechanisms studied and the resulting obser-
vations about BitTorrent performance. Many of the performance systems in three aspects. First, BitTorrent application does not
studies also suggested modifications to BitTorrent’s mechanisms provide a search function. Instead, peers are expected to peruse
to further improve its performance. We also present a survey of central-directory based search facilities provided by several
the suggested improvements and categorize them into different supporting websites. Second, it provides a file-level sharing
groups. policy instead of a directory-level sharing policy. Finally, it
Index Terms—BitTorrent, Large file downloading, Peer-to-peer employs a bartering mechanism such as TFT among peers [6],
Networks. such that unless a peer contributes to the ongoing downloading
of file pieces, it will not be able to download the file.
I. I NTRODUCTION The widespread popularity of BitTorrent has attracted the

T HE PEER-TO-PEER (P2P) paradigm has proved to be

a very effective approach for designing scalable and
robust networking applications. This approach overcomes 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
scalability limitations of the traditional client/server approach. performance. These studies affirmed the importance of TFT
Large scale file-sharing and file-distribution applications were and rarest first mechanisms for BitTorrent’s success. At the
the first to exploit the new paradigm. BitTorrent [1], one of same time, several of these studies identified the impact of
the most popular peer-to-peer (P2P) file distribution applica- BitTorrent’s phenomenal success on the traffic in computer
tions, was introduced by Bram Cohen in 2001 [2]. BitTorrent networks. For example, one of the significant problems is the
was quickly embraced by several content providers such as increase in file downloading traffic in the Internet. Reports
Lindows, Blizzard and most Linux distributions as a scalable show that BitTorrent has become the largest source of P2P
way of delivering content, decreasing the load on congested file sharing traffic in the Internet: BitTorrent traffic accounted
servers, minimizing the distribution cost and reducing the for 18% of broadband traffic in 2006 [7]. Karagiannis et al.
downloading time for users. BitTorrent is a second generation [8] also report that BitTorrent occupies about 13-15% of the
P2P file sharing system [3], which focuses on organizing the access link bandwidth at a residential university. Furthermore,
peers that are interested in downloading the same file into an BitTorrent presents significant traffic-engineering challenges
overlay network, called a Torrent. Interestingly, as more users for Internet service providers (ISPs). Current implementations
join a Torrent, the downloading rate that is achieved by all of BitTorrent ignore the underlying Internet topology or ISP
the peers increases. BitTorrent is noted for introducing several link costs and result in a large amount of cross-ISP traffic
innovative mechanisms including tit-for-tat (TFT) and rarest [9]. These issues show that BitTorrent protocol has further
first (RF) that it uses for cooperative file distribution among scope for improvement in its core mechanisms. For example,
the participating peers. one of the most criticized of BitTorrent’s mechanisms is the
Any file distribution application implemented using the P2P random neighbor selection strategy, which prolongs the file
paradigm requires two different functions to be supported: (1) downloading time and results in inefficient usage of network
a search function that enables peers to locate the content that resources. Some researchers also suggested BitTorrent-like
they are interested in among the participating peers, and (2) protocols [10], [11], [12], which modify the BitTorrent mech-
anisms to further improve performance.
Manuscript received 17 June 2008; revised 14 December 2008. The work
described in this paper has been supported by HK RGC under RGC- In this paper, we present a comprehensive survey of all
Competitive Earmarked Research Grant HKUST 617907 the performance studies on BitTorrent. We summarize the
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 salient findings of these performance studies. We then present
Kong (e-mail: [email protected]). a survey of all the improvements that have been suggested
J. K. Muppala is with the Dept. of Computer Science and Engineering, for BitTorrent and its mechanisms by various researchers. We
The Hong Kong University of Science and Technology, Clear Water Bay,
Kowloon, Hong Kong (e-mail: [email protected]). categorize these improvements into various groups so as to
Digital Object Identifier 10.1109/SURV.2010.021110.00036 enable us to understand the interplay among the suggested
1553-877X/10/$25.00 
c 2010 IEEE
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 141

improvements. Surveys of P2P networks in general, like Web Server Tracker

[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. 2 3
The remainder of this paper is organized as follows. In 1

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 5
improve its performance. Then we compare the suggested im- 4
Peer C
provements and categorize them into several groups. Finally,
we conclude this paper in section V. Peer A [Seed]
[Leecher]

II. OVERVIEW OF B IT T ORRENT

Peer B
A. What is BitTorrent [Leecher]
As stated earlier BitTorrent (BT) is one of the most popular
applications for distributing large size files. The application Fig. 1. BitTorrent file sharing process
is implemented as a hybrid P2P system, with most of the
and obtain a new list. The data flow between two connected
interaction directly among the peers, but requiring occasional
peers is bidirectional over the connection.
interaction with a server for locating peers. The BitTorrent
The steps in file sharing using BitTorrent are shown in
protocol requires peers to organize themselves into an overlay
Fig. 1. There are a total of five steps:
network, with connections among the peers, for each file being
distributed. This overlay network is called a Torrent. Each file 1) When peer A wants to join a torrent, it first downloads
being distributed by BitTorrent requires the establishment of the corresponding .torrent file from a web server.
separate torrent. Besides the peers, a tracker and a web server 2) Then, peer A contacts the tracker and asks for a list of
play an important part in file distribution using BitTorrent. peers that are already participating in the torrent.
The tracker is a special infrastructure node which stores 3) Next, the tracker sends a list of about 50 peers which
meta-information about the peers that are currently active are participating in the torrent.
within a torrent. Peers interact with the tracker using a simple 4) Then, peer A selects 35 peers from the list as its
protocol layered on top of HTTP in which a peer sends neighbors, and establishes connections with them.
information about the file it is downloading and the port 5) After the connections are established, peer A can ex-
number [2] to the tracker. The tracker does not participate in change file pieces with the neighbours.
the actual distribution of the file, but only serves the purpose The file exchange among peers uses a swarming technique
of enabling peers to find each other. [18], in which the file is broken into fixed size pieces (typically
The peers that are part of a torrent can be classified into 256 KB each), and peers exchange pieces with each another.
two types: a seed and a leecher. A seed is a client that has a When a piece is completely downloaded, its SHA1 hash
complete copy of the file and remains in the torrent to serve is computed and compared with the value in the metafile
other peers. For a torrent to get started, we need at least one (.torrent). If the values match, the peer will announce the
initial seed that provides the entire content for download. A availability of the complete piece to its neighbours.
leecher is a client that is still downloading the file. BitTorrent uses pipelining in order to keep its TCP con-
When a file is to be made available for distribution through nections operating at full capacity. To facilitate this, a piece
a torrent, then a special file with a .torrent extension is is divided into sub-pieces, typically of 16 KB in size, which
made available on a web server. The .torrent file contains are called blocks or chunks. Requests for sub-pieces are kept
information about the file including its length, name, hashing pending in a pipeline, typically five in number. Each sub-piece
information, and the url of a tracker [2]. A peer that wants arrival will trigger a request for another sub-piece. This way
to download a file first obtains the corresponding .torrent the connection is kept busy most of the time.
file from the web server. Then, the peer contacts the tracker A peer downloads pieces not only from the seed(s), but
and requests a list of IP/port pairs of other peers that are also from other peers, thereby substantially reducing the
already participating in the torrent. When the tracker receives load on the seed(s). A peer usually can serve four peers
a request, it will send the requesting peer a list of potential simultaneously, and it chooses the best four peers to unchoke
neighboring peers, usually 50, that are selected randomly from and chokes other requesting peers. Choking is a temporary
the set of all active peers within the torrent. After obtaining the refusal to upload, but the connections are not closed.
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 B. Mechanisms of BitTorrent
for the joining peer. If the number of neighbors connected to The mechanisms employed by BitTorrent mainly consist
a peer falls below 20, the peer will again contact the tracker of peer and piece selection strategies. A good peer selection
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 142 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010

A’s Peer Set

• Tit-for-tat: The tit-for-tat strategy is used by a peer to
preferentially upload to its neighbors who provide it
Interested Set the best downloading rates. At any given time, a peer
exchanges file pieces with a fixed number of neighboring
Mutually Interested Set peers, usually at most four. A leecher preferentially
uploads to the best three neighbors who provide it the best
Regular downloading rate and chokes others. Every 10 seconds,
Unchoke the leecher reevaluates the downloading rate from all
Set
Optimistic peers who are sending data to it. If another peer offers
Unchoke better downloading rate, the leecher will choke the peer
Set
with the smallest downloading rate among the current
peers it is serving, and unchoke the better peer. This
Fig. 2. Peer Sets
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,
strategy should maximize the service capacity of the system, BitTorrent incorporates an optimistic unchoking policy.
and an efficient piece selection strategy should guarantee that Every 30 seconds, a peer unchokes a neighboring peer
each peer can find pieces that are interesting to it from its randomly regardless of its uploading rate. The benefit of
neighbors. A detailed description of these mechanisms is Optimistic Unchoking is that it enables a peer to discover
given in [2]. Here we briefly summarize these mechanisms better neighboring peers to exchange file pieces with
for completeness so that further discussions in the subsequent since other neighbors may have higher capacity for up-
sections can be placed in proper context. loading than the currently unchoked peers. If a peer uses
As already mentioned in the previous section, the peer set only the TFT mechanism, there will be no opportunity for
of a peer A is the set of all the neighbors of A. Peer A will discovering other peers that can provide higher uploading
consider a peer B belonging to its peer set as an interested rate. In addition, this strategy is especially useful for
peer, if Peer B has pieces that A does not possess. We can newly joined peers to get started.
define an Interested Set of Peer A as the set of interested • Anti-snubbing: A peer may be choked by the peers it
peers, which is a subset of its peer set. A peer can continue was formerly downloading from, thereby getting poor
its download only if its interested set is non-null. Ideally Peer downloading rates. To address this problem, when a peer
A should use appropriate mechanisms to maximize the size notices that some time has elapsed without getting a
of its interested set. Doing so increases the ability of Peer single piece from a neighboring peer, the leecher assumes
A to download file pieces from the maximum number of it is ’snubbed’ by that peer and does not upload to it any
other peers. Similarly, Peer A considers peer B as a mutually further through regular unchoke.
interested peer if and only if peer B is in peer A’s interested • Upload Only: Once a peer finishes downloading the entire
set and peer A is also an interested peer for Peer B, i.e., Peer file, it becomes a seed. Since seeds have nothing to
A has pieces that Peer B does not possess. It is beneficial to download, they cannot select peers based on downloading
maximize the size of the mutually interested set because this rates. Instead, the seeds prefer to upload to peers with
increases the chances for trading pieces with the neighbors, better uploading rates [2]
thus effectively increasing the downloading rate. 2) Piece Selection Strategy: The second major strategy that
A peer A considers a peer B as a member of its regular plays an important role in the success of BitTorrent is the piece
unchoke set if peer B is unchoked by Peer A using the selection strategy. When a peer wants to download pieces from
TFT mechanism. Peer A considers peer B as a member of its neighbors, it adopts several piece selection strategies, which
its optimistic unchoke set if peer B is unchoked by Peer A includes the following four mechanisms:
due to the optimistic unchoking mechanism. Details of these
• Strict Priority: In BitTorrent, peers concentrate on down-
mechanisms are given later in this section. We illustrate these
loading a whole piece before requesting another piece.
sets and their relationships further in Fig. 2. These sets play
Thus if a sub-piece is requested, then subsequent sub-
an important role in determining the downloading experience
pieces of the same piece will be requested preferentially
of any peer.
in order to complete the download of the whole piece as
1) Peer Selection Strategy: The peer selection strategy uses soon as possible, because only complete pieces can be
four main mechanisms: tit-for-tat (TFT), optimistic unchoking traded with others.
(OU), anti-snubbing, and upload only. Summed up together, • Rarest First: Peers often prefer to download pieces which
they form the basis for the choking algorithm used by a are the rarest among their neighbors [2]. This strategy is
peer. The major aim of these mechanisms is to improve the called the rarest first algorithm which works as follows.
downloading experience of those peers that contribute to the Each peer maintains a list of the pieces that each of its
file exchange by uploading pieces to others, and punish free- neighbors possess. This list is updated each time a copy
riders, i.e., those peers that download pieces from others of a piece becomes available from its neighbors. It then
without uploading any pieces. constructs a rarest-pieces set which is the list of those
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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 but also theoretically optimal downloading time for peers. In
neighbors. It then selects the piece to download that is FOX, the authors assumed that all peers are greedy. In the
rarest among its neighbors. This strategy increases the overlay structure of FOX, there are N clients downloading
chance of a peer trading with its neighbors as it has pieces a file from one server. Suppose Ri is the peer connected
that others want. This also benefits the seeds by reducing to the server, and Ti is the subtree rooted at Ri , and Li
their server burden, since they need to serve out less is the set of leaves of Ti . To get a quick downloading
copies of the same piece. Furthermore, the departure of time, FOX fills the available outgoing link at the server, in
a seed will not leave the remaining downloaders without which each Ri requests a different block from the server,
the opportunity to exchange file pieces, because this and each leaf only sends blocks to leaves which reciprocate
strategy ensures that all the pieces are quickly distributed the gesture. Peers in FOX have two properties. The first
to at least some of the leechers [2]. property is the give-and-take symmetry, which means each
• Random First Piece: There is an exception to the rarest of peer’s outgoing edges has a corresponding incoming edge,
first policy when a peer first joins a torrent. Since the peer instead of allowing peers to arbitrarily match up. The second
does not have any pieces, so it would like to download a property is that the connections are stable. Based on these
whole piece quickly so that it can get ready to reciprocate two properties, peers’ misbehavior, such as receiving but not
with the TFT algorithm. In this case, the new peer will sending, can be detected. In addition, FOX can deal with the
not seek the rarest piece. Instead, it will download the last-block collapse, in which the finished nodes quit the system
first piece randomly in order to make sure it can get a leaving the remaining nodes without any help. Fox uses a
whole piece as fast as possible. cryptographic method, in which the final blocks are encrypted
• Endgame Mode: This mode is adopted by a peer towards by the internal nodes before sending to the leaves. Then, the
the end of downloading the file. If a piece is requested leaves use their own keys to encrypt the encrypted blocks.
from a peer that has a slow transfer rate, the downloading So both nodes and leaves need others’ keys before decrypting
time will be prolonged. To address this problem, a peer the encrypted blocks and completion of download. By using
requests all of its neighbors for blocks it has not received. this approach, both internal nodes and leaves hold information
Once a block is obtained, the peer cancels the request for that the others need to complete their download. So FOX can
that block from its neighbors so as to decrease bandwidth make all nodes complete their download at the same time and
wastage due to redundant downloads. avoid the last-block collapse.
In P2P systems, users often decide which block to download
without considering whether the block has already been sent
C. BitTorrent-like Systems
out by the server at least once or not. This causes subop-
Some researchers proposed BitTorrent-like systems such timal distribution decisions. Gkantsidis et al. [10] proposed
as Slurpie [12], Fox [11], and Avalanche [10]. They have Avalanche using networking coding technique. Like BitTor-
a lot of similarities in their operations with BitTorrent. The rent, nodes join the system by contacting a centralized server
common goal of these approaches is to improve the file sharing which provides a random subset of other users. Nodes also
performance. employ rarest first algorithm to decide which block to transfer.
Slurpie [12] is designed with the goal of reducing the In Avalanche, there are two incentive mechanisms. The first
clients’ downloading time and making the client sets scalable. one is giving priority to exchange over free uploading to
In Slurpie, when a node wants to download a file, it first other nodes. The second one is that nodes do not upload
contacts a server called a topology server. The topology server to users unless they have also received sufficient content in
sends the node a list of other nodes which are downloading return, which is similar to the TFT policy. The principle of
the same file. The node then contacts the other nodes in the network coding is to allow nodes to encode packets. When a
list to form a mesh. Similarly, the file is divided into fixed client wants to send a packet to others, it will make a linear
size blocks, and is exchanged through the mesh. The authors combination of all the available information and then send
claim that Slurpie has advantages in adapting to varying it out. At the receiver’s side, the original information will
bandwidth and achieves better scaling with the number of be reconstructed after receiving enough linear combination
neighbors compared to BitTorrent. Slurpie employs an effec- of packets. With this method, the block propagation becomes
tive downloading bandwidth estimation technique described more efficient and the system is more robust under extreme
in [12] which reports three different states of bandwidth for situations like the sudden departure of nodes.
each node: underutilized, at-capacity, and throttle-back. Each The characteristics of the above BitTorrent-like systems are
node in Slurpie maintains a target number of neighbors, and summarized in Table I. As can be seen from the table, the
this number is continually updated according to the available above three systems not only share some mechanisms in com-
bandwidth. The bandwidth estimation algorithm runs every mon with BitTorrent, but also have their own characteristics.
second, and if the node’s state is underutilized and the node Slurpie is very close to BitTorrent in its operation. A major
needs more blocks, a new connection will be opened. By using difference is that it provides bandwidth estimation technique
these strategies, Slurpie can improve the system performance which enables it to scale well as the group size increases. FOX
as the size of the network increases. proposes a fair protocol to provide an optimal downloading
Levin et al. [11] proposed a Fair, Optimal eXchange proto- time. While Avalanche is very similar to BitTorrent, it uses
col, called FOX to achieve a fair file swarming. FOX provides network coding technique to make the scheduling of the
not only an effective application of the TFT mechanism, content propagation easier.
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 144 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010

TABLE I
B IT T ORRENT- LIKE S YSTEMS

BitTorrent-like system Special Characteristics Common characteristics with BT References

Slurpie Bandwidth estimation technique. The number of neigh- Central server. Mesh formed [12]
bors is updated according to the available bandwidth among peers.
FOX Peer connections are based on give-and-take symmetry. TFT policy [11]
The connections are stable.
Avalanche Each piece of the shared file is encoded by the nodes Central server. Rarest first policy. [10]
using network coding. Incentive mechanism.

III. P ERFORMANCE OF B IT T ORRENT • The uploading and downloading rates are correlated,
The popularity of BitTorrent has spurred a large number which means that the TFT policy works well.
of studies of its performance. Some measure its behavior on We note that the perspective presented in this paper is based
real torrents; others use the simulation to observe its perfor- on the detailed analysis of a single torrent, which may be
mance while others use mathematical analysis-based models its limitation in that the perspective is limited. However
to evaluate its performance. In this section, we first organize the observations are still relevant and corroborated by other
the papers based on the performance evaluation method used. studies reviewed in this section later.
The main purpose of organizing this section according to the Pouwelse et al. [6] presented a measurement based study
performance evaluation methods used is to group together of BitTorrent’s availability, integrity, flash crowd handling, and
the papers using similar approaches to evaluation. We do download performance based on measurement data obtained
notice that each evaluation technique lends itself to examining from June 2003 to March 2004. They also measured Suprnova,
certain aspects of BitTorrent performance better than the other one of the most popular websites for BitTorrent clients to find
techniques. We highlight the advantages and shortcomings of files. Their experiment consisted of two parts. They monitored
each approach in the respective section. In the final subsection the global BitTorrent components using three scripts: a Mirror
we concentrate more on reviewing the techniques from the script for measuring the Suprnova mirrors’ availability and
perspective of the issues addressed and the different aspects response time, a HTML script for gathering and parsing
of performance studied. Suprnova mirrors’ HTML pages and downloading new files
with the extension of .torrent, and a Tracker script used to
A. Measurement parse the .torrent files for new trackers and check the trackers’
Several measurement-based studies of BitTorrent perfor- status. They also observed the actual peers in the system using
mance have been presented in the literature. Most of the scripts: the Hunt script was used to select a file to follow and
measurement-based experiments lasted for several months. initiate a measurement of all the peers who are downloading
Generally, the following five methods were used: (a) ana- that file, the Getpeer script was used to obtain the IP addresses
lyzing the tracker logs obtained from the trackers, (b) using of peers downloading the file from the tracker, the Peering
scripts to gather information from torrent websites and directly script for measuring the download progress and uptime of
from peers, (c) analyzing packet traces collected at Internet peers in parallel. They found that:
access link, (d) joining an ongoing torrent with a modified • The number of users is related with the number of seeds
client specially instrumented to collect event logs, and (e) in BitTorrent.
conducting experiments on network testbeds like PlanetLab • Most users often stay in the system to be seeds for several
or user-constructed networks of PCs. Observing the tracker hours after finishing their downloading.
log can enable us to get the information on the global view • BitTorrent can deal with flashcrowds efficiently.
of BitTorrent performance; whereas using a modified client • Fewer number of seeds does not mean that the lifetime
enable us to observe the individual behavior of peers. of the file is short. Even files with only one seed can still
Izal et al. [18] evaluated the performance of BitTorrent have a long content lifetime.
mechanisms based on five months of measurement data The authors concentrated mainly on the user behavior and the
collected from BitTorrent peers. This was one of the first content distributed by BitTorrent. This gives a good macro
comprehensive evaluation of BitTorrent using measurement. view of BitTorrent behavior. However this measurement does
First, they used the tracker log of Linux RedHat 9 distribution not directly shed light on the core mechanisms in BitTorrent
for their evaluation. They found that: and their effect on its performance.
• BitTorrent clients are altruistic. Legout et al. [22] conducted a detailed measurement study
• Seeds contribute more than twice the data uploaded by aimed at understanding the behavior of the rarest first al-
leechers. gorithm and the choking algorithm on real torrents. They
• The proportion of seeds is higher than 20%. instrumented a BitTorrent client and observed its performance
Then, they used an instrumented client to join the torrent to in several torrents. They observed the log of messages sent
observe the individual behavior of a peer. They observed that: or received, the log of each state change, the log of the
• After obtaining the first few pieces, peers begin trading rate estimation, and the log of important events such as end
pieces soon, which indicates that the rarest first policy game mode. Their results showed that (1) the rarest first
works well. algorithm prevents the reappearance of rare pieces and avoids
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 145

TABLE II
S UMMARY OF M EASUREMENT- BASED S TUDIES OF B IT T ORRENT P ERFORMANCE

Measurement Methods Results References

Observing the tracker log BitTorrent clients are altruistic. The rarest first policy works well. The [18]
TFT policy works well
BitTorrent can support large files. The incentive mechanisms are effec- [19]
tive
TFT is imperfect at preventing free-riding [20]
The distribution of performance among peers is roughly uniform. TFT [21]
is the main factor that affects the performance
Using Mirror script, HTML script, Tracker BitTorrent can deal with flash crowds effectively [6]
script, Hunt script, Getpeer script, and Peering
script
Analyzing packet traces collected at Internet BitTorrent is locality-unaware, causing the same content to be down- [8]
access link of a residential university network, loaded into the same locality (defined around an ISP) multiple times.
and Tracker Logs 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
Using a client on real torrent The rarest first algorithm prevents the reappearance of rare pieces and [22]
avoids the last pieces problem. The choke algorithm is fair and robust.
Peers behind a firewall have low bandwidth usage. Peers behind a [23]
firewall often disconnect early
Using a modified version of BT on PlanetLab The choking algorithm can cluster peers with similar bandwidth, reward [24]
contributing peers, and make peers have a high upload utilization
The initial stage is not predictive of the overall performance. The [25]
unchoked network is scale-free. There is no clustering except the initial
stage.
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 peer-assisted content delivery and its impact on ISPs, we
TFT with a bit level TFT solution does not necessarily yield are specifically concentrating on their measurement based
better fairness; (3) the choke algorithm is fair and robust. evaluation and observations about BitTorrent. They indicated
Their recommendation is that the rarest first strategy already that BitTorrent is locality-unaware which severely increases
achieves excellent performance and there is no compelling the ISPs’ cost, typically resulting in the same content being
need for more complex strategies. They do observe some scope downloaded into the same ISP multiple times from external
for improvement in the choke algorithm. While this study peers. They also found that users requesting the same content
sheds light on the important properties of the choke algorithm, have a 30%-70% overlap in their lifetimes, consequently
however their measurement is from a single-peer view point helping each other in the download. An additional 20%-40%
which cannot evaluate the dynamics of BitTorrent. Also, the improvement in terms of cross-ISP downloaded content can
reasons behind the observed properties of the choke algorithm be effected by giving incentives for users to stay around after
are unclear. they complete the download.
Bellissimo et al. [19] collected statistics from two trackers Andrade et al. [20] focused on measuring the cooperation in
to conduct their measurement-based evaluation. The statistics BitTorrent. In their measurement, they defined three metrics:
included the user’s random ID, how long the user had been (1) free-riding ratio, which is the percentage of free-rider
connected to the torrent, and the number of bytes uploaded, peers, (2) seeding ratio, which is the percentage of seeds in
downloaded, and remaining. Their results showed that (1) a torrent, and (3) sharing ratio, which is the total uploaded
BitTorrent can serve large files effectively; (2) the incentive data divided by the downloaded data. They evaluated the
mechanisms are effective. metrics by collecting logs in BitTorrent communities such
In [23], Skevik et al. focused on the effect of the firewalls as bt.etree.org, and piratebay.org. Then, they measured the
and the time the peers have stayed in the system. They downloading time for free-riders and non free-riders, and the
monitored the torrents of RedHat and Mandrake Linux for amount of free-riding and low-sharing behavior in BitTorrent
several months by executing a modified client and a crawler. communities: etree and easytree. Their results demonstrate
They observed that peers behind a firewall have low bandwidth that BitTorrent’s TFT protocol does discourage free-riding
usage and they are more likely to leave the system early successfully. However, if there are a large number of seeds
because of their poor download performance. in the torrent, the TFT mechanism may not work effectively
Karagiannis et al. [8] quantified the influence of BitTorrent to prevent free-riding.
on ISPs using two types of real measurement data, the Legout et al. [24] conducted experiments on private torrents
BitTorrent tracker log for RedHat v9.0 and the payload packet and collected data from peers in a controlled environment.
traces collected using a monitor installed on the access link They focused on the choking algorithm in BitTorrent to
of the network of a residential university in USA with 20,000 observe peers’ individual behavior during the downloading
users. While the paper considers the more general issues of process. Their experiments were performed on PlanetLab.
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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 much greater flexibility to vary the various mechanisms used
changes using a modified version of BitTorrent. Their results in BitTorrent and observe the impact through extensive exper-
reveal that the choking algorithm can cluster peers with similar imentation. Compared with the measurement-based approach,
bandwidth, reward contributing peers efficiently, and make the simulation-based approach has two advantages. One ad-
peers have a high uploading utilization. vantage is the greater flexibility in controlling the various
Dale et al. [25] looked at the topologies of different kinds configuration parameters of BitTorrent mechanisms. Another
of networks that can be clearly identified within a BitTor- advantage is that it allows us to study the impact of variations
rent swarm. In particular they investigated the evolution of in a particular mechanism while keeping the rest fixed, which
four kinds of networks that can be used to characterize a is very difficult to achieve in measurement-based experiments.
swarm: (1) Connection Network: the network of neighbors Simulation-based approaches are especially useful in topology
that each peer maintains. (2) Interest Network: the network related studies since a crawler does not yield information
defined by the interest that peers maintain in other peers. about a peer’s neighbors in BitTorrent because of the lack of
(3) Unchoked Network: the network that is formed by the distributed mechanisms for peer discovery or lookup. Further-
peers and the neighbors that are unchoked by each peer. (4) more a peer’s connectivity information is not shared with the
Download Network: the network that is formed linking peers tracker and hence tracker logs do not yield overlay information
to other peers from which they are downloading. Among [31].
these networks, all networks are directed graphs except for Many of the simulation based studies make simplifying
the connection network which can be considered undirected. assumptions for the sake of ease of simulation. Sometimes,
In their experiments, they used a modified BitTornado [27] these assumptions may not necessarily be realistic. Very often
client running on more than 400 nodes of PlanetLab that form the underlying network topology is simplified and packet-level
the swarm. They found that: dynamics are not fully represented. Most of the simulators
• The initial stage is not predictive of the overall perfor- concentrate only on representing the behavior at the overlay
mance. network level. Thus the impact of the underlying physical
• The unchoked network is scale-free. network is not clearly reflected in the simulations. Also many
• There is no evidence of clustering in their experiment simulations consider a smaller set of peers in a swarn than is
except in the initial stage. normally encountered in practice. The results from simulation
Rasti et al. [21] pointed out that several of the measurement studies are summarized in Table III.
studies of BitTorrent were just based on observing the behavior Felber et al. [28] presented a simulation based study of
of a few instrumented peers. This may not necessarily provide the effect of different peer and piece selection strategies on
a representative view of the overall BitTorrent performance. the performance of BitTorrent. In their peer selection strategy,
Instead, a distribution of the performance indices among the they assumed that a peer has a global view of the state of
peers provides a broader view of BitTorrent performance. all the other peers and thus selected the most suitable peer
Furthermore, the effect of different peer-level and group-level to exchange pieces with based on the peer selection strategy.
properties on the performance needs to be identified. To verify They considered several peer selection strategies including
their theory, they analyzed a BitTorrent tracker log to estimate random, and several variations based on the number of pieces
the distribution among peers in a torrent and identify the effect possessed/missing in the other peers. Their results demonstrate
of the peer-level and group-level properties on the perfor- that there is no clear winner among all the strategies in
mance. They found: (1) the distribution of performance among all scenarios. Similarly, they considered two different piece
peers is roughly uniform; (2) the TFT mechanism is the main selection strategies: random and rarest. Among these the rarest
factor that affects the performance. They could not observe strategy seems to deliver the most consistent performance
any direct relationship between peer-level performance and across different peer selection strategies.
Bharambe et al. [29] evaluated the impact of BitTorrent’s
main peer and group-level properties. Furthermore, the average
core mechanisms on overall system performance under a
upload rate of individual peers had the highest correlation with
variety of flash-crowd workloads using a simulator, which
its observed performance [21].
can model not only peer’s activity such as joins, departures,
Lei et al. [26] did an experiment to test the performance
and block exchanges, but also BitTorrent’s mechanisms like
of BitTorrent’s neighbor selection. Their experiment includes
TFT and rarest first. To make the model simple, they did not
96 PCs, each of the same configuration. They used BitComet
model network propagation, the packet-level dynamics of TCP
as the BitTorrent client for their experiments. They observed
connections and the shared bottleneck links in the interior of
that the random neighbor selection in BitTorrent does not take
the network. They focused on the link utilization, the fairness
into account the communication cost, which will result in low
and the optimality to quantify the effectiveness of BitTorrent.
transmission rate and high operating cost.
They made their simulator named BTSim publicly available
We summarize the various observations on BitTorrent per-
[34]. From the simulation results they found that:
formance obtained through measurements in Table II.
• BitTorrent is robust and scalable, and ensures high uplink
bandwidth utilization.
B. Simulation • The TFT policy fails to prevent the unfairness in BitTor-
Evaluating the impact of various factors on the overall rent.
performance is very difficult using measurement-based tech- • The rarest first policy is critical in ensuring that the newly
niques. Simulation based studies on the other hand provide joined peers have something to exchange with others.
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 147

TABLE III
S UMMARY OF S IMULATION - BASED S TUDIES OF B IT T ORRENT P ERFORMANCE

Simulation Features, Assumptions and Shortcomings Results References

Event-driven simulator with resolution of a millisecond. No appreciable impact of peer selection strategies. Rarest first [28]
Peers have global view for peer selection piece selection strategy gives balanced performance
The event-driven simulator models both peer’s activ- BT is robust and scalable at ensuring high uplink bandwidth [29]
ity and BT’s mechanisms. Network propagation, the utilization. The TFT policy fails to prevent unfairness in Bit-
packet-level dynamics of TCP connection and the Torrent. The rarest first policy is critical to ensure the new joined
shared bottleneck links are not modeled peers have something to exchange with others.
Simple simulator contains one seed serving a file com- By selectively uploading to its poorest neighbor, the peer [30]
posed of n pieces. Network latency, connection costs ensures that the data remain in the network for the longest
or complex topologies not modeled. period, thus reducing its own burden.
The simulator is developed in MATLAB The authors The overlay of BT is not a random graph. A large number of [31]
simulated the tracker protocol in mainline client 4.0.2. NATed peers decreases the robustness of the overlay to attacks
significantly.
The simulator runs in rounds and each round lasts for The peer set size (PS) and the percentage of outgoing con- [32]
10 seconds A leecher sorts peers based on the data they nections (OC) have a significant impact on the BitTorrent’s
offer The bottlenecks is downlinks/uplinks. The authors performance. Decreasing OC is more efficient than increasing
did not model any delay. PS. The choice of PS and OC can have an impact on the overlay
structure.
The set of nodes involved in BT session is fixed. The current BT is far from the optimal solution that minimize [33]
Endgame mode is not simulated. The neighbor set of the end-to-end delay of distributing a file from a source to
each BT node is static. Delays are not simulated except multiple receivers. The peer selection strategy may lead to
network packet transmission delay. mismatched peering and make BT deviate more from optimum.

The simulations were restricted to a peer set no larger than 15 the diameter of the overlay, the robustness of the overlay to
peers, while the real implementations can have up to 80. This attacks and high churn rate. They pointed out that a large peer
might influence some of the results since the accuracy of the set size can make BitTorrent more efficient. More importantly,
piece selection strategy is affected by the peer set size. they showed for the first time that the overlay in BitTorrent is
Adar [30] demonstrated through simple simulations that a not a random graph, and connectivity of a peer to its neighbors
peer possessing most of the file pieces can significantly reduce depends on its arriving order. In addition, they showed that
its effort in servicing other peers while at the same time a large number of NATed peers (peers behind a NAT or a
improving network performance. They simulated the effect of firewall) significantly decreases the robustness of the overlay
bit welfare. There is only one seed serving a file composed to attacks.
of n pieces in the simulator and three seed strategies are Wu et al. [33] were specifically interested in answering the
considered: Random, in which the seed chooses the peer to question of how close BitTorrent is to the optimal solution.
give piece randomly, Preferred, in which a peer is chosen to They proposed a centrally scheduled file distribution (CSFD)
receive pieces from a preferred subset, Poorest-first, in which protocol that can minimize the total elapsed time. In addition,
the peer who has the fewest pieces is chosen. Based on their they compared the performance of BitTorrent and CSFD. They
experiments, they found that the Poorest-first runs better than found that the current BitTorrent is far from the optimal
other algorithms. The experimental results showed that by solution that minimizes the end-to-end delay of distributing
selectively uploading to its poorest neighbor, the peer ensures a file from a source to multiple receivers. Moreover, the
that the data remain in the network for the longest period, thus peer selection strategy of BitTorrent may lead to mismatched
reducing its own burden. peering, and make BitTorrent deviate from the optimum.
Urvoy-Keller et al. [32] adopted a simulation approach
to evaluate the impact of the overlay topology parameters C. Analytical Modeling
on the BitTorrent performance. Their results show that the Several analytical models of BitTorrent-like file sharing
two parameters, the peer set size (PS) and the percentage of system have been proposed in the literature (See Table IV).
outgoing connections (OC), have a significant impact on the Existing analytical studies can be classified into two cate-
BitTorrent’s performance. In addition, decreasing OC is more gories: homogenous, where all peers have the same upload
efficient than increasing PS. Furthermore, the overlay structure and downloading rate; and heterogeneous, where peers have
can be affected by the choice of PS and OC. different uploading and downloading rates. Analytical models
Al Hamra et al. [31] followed up on the work by Urvoy- are clearly able to reflect the effects of different parameters
Keller et al. [32]. They specifically concentrated on evaluat- on BitTorrent performance. They permit efficient and detailed
ing the properties of the distribution overlay of BitTorrent. exploration of the parameter space to evaluate the effect of
They presented an in-depth study of the overlay topologies variations of not just a single parameter, but also the combined
in BitTorrent by conducting simulation experiments using a effect of variations of several parameters. However, many of
simulator developed in MATLAB. They analyzed the relation these models are based on sometimes unrealistic assumptions
between the overlay properties and the BitTorrent performance like peers having global information about the state of all
considering the following four parameters: average peer set peers, simplifying assumptions on the underlying network
size, the time for a peer to reach its maximum peer set size, topology, and on the arrivals and departures of peers. Most of
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 148 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010

TABLE IV
S UMMARY OF A NALYSIS - BASED S TUDIES OF B IT T ORRENT P ERFORMANCE

Focus of Analysis Modeling Results References

Approach
Service capacity and fairness Homogenous P2P systems achieve favorable scaling in terms [35]
Branching process, of average download delay with increasing load.
Markovian ε-peerwise fairness is a practical approach, for
example as used in BitTorrent.
Scalability and incentive mechanisms Homogenous The system scales well with the number of peers. [3]
Fluid-Flow The incentive mechanisms are efficient. Optimistic
unchoking may encourage free-riders
Single-Torrent system, file availability Homogenous BitTorrent is good at dealing with “flash crowds.” [36]
Multi-torrent system performance Fluid-flow 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
The distribution of the individual chunk Homogenous The modified routing policies perform better [37]
Evaluation of file distribution time Heterogeneous Obtains an explicit expression for the minimum [38]
Application-level differentiated service Fluid-Flow achievable time to distribute a file.
model
Influences of free-riding Heterogenous BitTorrent mechanism can prevent free-riding suc- [39]
Fluid-flow cessfully
The distribution of the peers’ download Homogenous Peer distribution in terms of their current download [40]
state, file availability and system death Fluid-flow, state follows a U-shaped curve, with most peers in
Markovian 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.
The download behavior and incentive Homogenous First attempt at a queueing model for BitTorrent [41]
mechanism Queueing Model protocol
Altruism and Incentive Mechanisms. Fo- Homogenous Altruistic contribution by high-capacity peers seem [42]
cus on reciprocation and upload band- Probabilistic to cause the most benefit in download performance.
width sharing model TFT policy is not the major cause for good perfor-
mance
Effect of Bandwidth Heterogeneity on Heterogeneous Heterogeneous bandwidth can have a positive ef- [43]
download performance Fluid-flow fect on content propagation
Two-class problem. Focus on Service Heterogeneous System of differential equations admits a unique [44]
differentiation among the classes. Band- Fluid-flow stable equilibrium point. Whether the system of
width diversity problem differential equations has a stable state or not
depends on the initial conditions.
File downloading time File availability Heterogeneous Stochastic differential equation approach, yields [45]
Both Single class and Multiclass Fluid-flow closed form solutions for several performance mea-
sures including average download time, system
throughput, number of peers and seeds. Sensitivity
analysis of performance measures.
File dissemination in P2P networks, Min- Heterogeneous de- Distribution time of any natural random strategy as [46]
imum file distribution time terministic adopted by BitTorrent is proportional to the optimal
time
Heterogeneous peers and performance Heterogeneous Simple model yields accurate results, verified with [47]
using simple models using balance equa- deterministic flow simulation. Token based scheme, which can yield
tions. balance equations better performance for low-bandwidth peers by
giving higher weightage to their upload contribu-
tion.

the analytical models rely on a fluid-flow approach to charac- • c, the downloading bandwidth of a peer.
terize the system behavior. In particular, the models represent • θ, the rate at which peers abort their download
the evolution of the system by tracking the number of peers, • γ, the rate at which seeds leave the system, and
X(t), and the number of seeds, Y (t), either as a function of • η, the effectiveness of file sharing.
time, or in steady-state. The evolution is represented using
Some models keep track of the different stages of downloading
(stochastic) differential equations involving various system
that a peer may exist at any point of time. This typically
parameters. Typical parameters that characterize the system
involves dividing the peers, X(t), into different subsets,
evolution include:
X1 (t), X2 (t), etc. each representing the different stages of
• λ, the arrival rate of new requests. The interarrival time download. However, this increases the number of simultaneous
is usually assumed to be exponentially distributed. equations that need to be solved and hence the complexity of
• μ, the upload bandwidth of a peer. the model.
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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 Guo et al. [36] followed up the idea of [3]. First they
terms of their “service capacity” for both transient and steady- analyzed the file downloading trace files obtained from two
state regimes. They considered an abstract model of P2P dedicated tracker sites that host multiple torrents. They also
file sharing system, and focused on peers that are sharing a analyzed BitTorrent metafile downloading traces that were
single file. They concentrated on the file download process and collected from a large commercial server farm hosted by a
assumed that each peer knows the address and file availability major ISP and a large group of users connected to the Internet.
at other peers. Furthermore, they assumed that no peer leaves From the trace analysis, they concluded that the peer arrival
the system before finishing their downloading. They first rate to a torrent decreases exponentially. They studied the
presented transient analysis of the P2P system using both torrent lifespan, the duration from birth until the file becomes
deterministic and branching process approaches to determine unavailable. They derived the expressions for the torrent
the average download delay. They demonstrated that P2P lifespan and the downloading failure ratio, the ratio of peers
systems achieve favorable scaling in terms of the average that do not complete the download to the total peer population
download delay with increasing load. They also show that for a torrent. They extended the fluid-flow model in [3] using
parallel uploads from a peer, where the upload bandwidth is the model for the exponentially decreasing peer arrival rate
used by a peer to simultaneously upload to multiple peers, to construct the torrent evolution model. They found that
is effective even when peers are uncooperative. Then they both peers and seeds increase exponentially initially, but then
modeled the stationary regime using a Markovian approach. decrease exponentially at a slow rate. The seed departures
Furthermore, they studied the fairness issue in P2P systems by result in the torrent being unable to keep up with the peer
considering different notions of fairness for service allocation service demand, causing the torrent to eventually die. The
in the stationary regime. Achieving global proportional fair- random arrival and departures of peers and seeds causes the
ness and pairwise proportional fairness is not easy for robust performance to fluctuate significantly over the torrent lifetime,
implementation. They then proposed an ε-peerwise fairness especially for small torrents. Furthermore, they demonstrated
which is better suited for a dynamic environment with peers that the biased service policy of seeds towards peers with
joining and leaving the system. This is the strategy used in higher download bandwidth might result in these peers not
BitTorrent with its choking algorithms. contributing their fair share of upload capacity to the torrent.
They also proposed a graph based multi-torrent for analyzing
Qiu et al. [3] presented an analytical model of BitTorrent inter-torrent collaborations. Their analysis demonstrated the
based on a fluid flow model. Their work was influenced in feasibility of using multi-torrent collaboration effectively to
part by the earlier work of Yang et al. [35]. In developing their improve performance.
model, they made several assumptions: downloaders become In [37], Arthur et al. proposed a BitTorrent-like model
seeds with a probability that is determined by parameters like using which they analyzed the data disseminating protocols
the number of peers; downloaders can abort their downloading of BitTorrent and related P2P networks. In this model, they
after a certain time that is exponentially distributed, and ignored the TFT strategy and assumed that all nodes have the
each seed stays in the system for a random time which same bandwidth. In particular, they modeled the routing of
is also exponentially distributed. They developed a simple data blocks on a directed graph over discrete time steps. They
deterministic fluid-flow model to describe the evolution of also analyzed some of the multiple network topologies and
the number of leechers and the number of seeds, in terms routing algorithms.
of the arrival rates, the upload and download bandwidths, Kumar et al. [38] presented an analytical model of peer-
the abort rate of downloaders, the rate at which seeds leave assisted file distribution time in P2P networks using a fluid-
the system and the effectiveness of file sharing. They used flow based model. Their aim is to study the advantages of
this model and studied the steady-state performance, with a peer-assisted file distribution system vis-à-vis a traditional
special attention to the protocol’s scalability and efficiency. client-server distribution system, specifically the scalability
They found that the average downloading time is not related and efficiency of the new approach. In particular, they were
to the peer arrival rate, and hence the system scales very well interested in finding the minimum distribution time for the
with the number of peers. Their model clearly demonstrated file to be distributed from the seeds to all the leechers in a
that the average downloading time decreases with efficient file general heterogeneous peer-assisted file distribution system.
sharing and increases with increasing seed departure rate, as In their model, they made two assumptions: The bandwidth
expected. Also changes in the uploading and downloading bottlenecks are in the uploading and downloading link rates
bandwidth impact the downloading delay to some extent. instead of in the Internet core, and every node participates in
They then presented a game-theoretic analysis of both the the file sharing until it completes the file download. Using
choke and rarest first algorithms. They show that the incentive this model they obtained an explicit expression for the min-
mechanisms in BitTorrent are efficient. They also show clearly imum distribution time. This expression explicitly shows the
that the optimistic unchoking might lead a free-rider to get relationship of the distribution time to the file size, the seeds’
upto 20% (In general, 1/(nu + 1)%, where nu is the number uploading rate, and the leechers’ uploading and downloading
of regular unchokes) of the maximum downloading rate. rate. They also point out that although their approach makes
They also presented experimental results from simulations to an unrealistic assumption about the flow of bits through a
validate the results from their models. The assumption about node compared to the chunk-based models that were proposed
global knowledge of all peers for peer selection is a major earlier, the resulting expressions provide a very good (lower-
limitation of this approach. bound) approximation for the file distribution time. However,
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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 heterogeneity on file transfer dynamics and content diffusion
by their approach. process in detail. In particular, they compared the performance
In [39], Yu et al. focused on analyzing the influence of free- of heterogeneous networks and the equivalent homogeneous
riding in BitTorrent System. They extended the model in [3] networks under different conditions of equivalence. Their
with two classes of peers: nonfree-riders and free-riders. Their results show that heterogeneous bandwidth can have a positive
results show that BitTorrent’s mechanism can guard against effect on content propagation among peers if appropriate
free-riding successfully. In addition, they discussed the system criteria were chosen.
death induced by free-riding and pointed out that retaining a Clévenot-Perronnin et al. [44] argued that the model in [3]
copy of each piece in the system can prevent system death. considers only homogeneous peers, but in practice, peers have
Tian et al. [40] extended the fluid-flow model in [3] to diverse bandwidth characteristics, such as Ethernet access,
study the distribution of the peers in different states of the dial-up modem access and broadband access. This limits
downloading completion. Towards this end, they explicitly the model’s applicability. Therefore, they presented a general
represent the peers as belonging to one of several states, multiclass model for heterogeneous peers with different ac-
S0 through SN , where 1/N is the level of granularity at cess bandwidth and multiple differentiated service classes. In
which they view the current download progress. The peer particular they considered in detail a system with two classes
state behavior is now viewed as a continuous time Markov of peers, distinguished by their different upload and download
chain. From this they derived the distribution of the peers in speeds. In their model, they allow an uploader (either a seed
each state of the download. The distribution is observed to or peer) to statically allocate its upload bandwidth to the
be a U-shaped curve. From the model, they found that seeds’ two classes of peers. They used this model to analyze two
departure rate and leechers’ abort rate can influence the peer problems: service differentiation and bandwidth diversity. For
distribution. Furthermore, the TFT strategy cannot improve the service differentiation problem, they described that the
the file availability or prevent the system death caused by a system of differential equations admits a unique stable equi-
sudden departure of a finished peer. librium. For the bandwidth diversity problem, they indicated
Saraswat et al. [41] presented a queuing model for BitTor- that whether the system of differential equations has a stable
rent using the Java Modeling Tools [48]. In their model, each state or not depend on the initial conditions.
BitTorrent client is modeled as a load dependent station. The Fan et al. [45] extended the model in [3] to obtain a
stations’ service request time corresponds to the time taken to fluid-flow model based on the stochastic differential equation
download a piece, the service time correspond to the inverse approach. In this model, they divided peers into three types:
of the downloading bandwidth, and the number of jobs in leechers that have a few chunks, leechers that have most
the queue represents the total number of pieces downloaded. chunks, and seeders. They define the connection probability
They set the service time dependent upon the number of jobs ρ(≤ 1) as the probability that a peer maintains connectivity
in the station to ensure that the incentive mechanisms work with another peer in the torrent. They analyzed the file
well. They focused on the download behavior and incentive downloading time and file availability in BT-like systems
mechanisms. with this model. They obtained closed-form solutions for the
Piatek et al. [42] presented a model for the altruism in average number of seeds and leechers, the average down-
BitTorrent. They made the following assumptions: (1) The loading time and the steady state throughput of the system.
distribution of a typical swarm may not be identical; (2) The The closed-form solutions show the effect of various system
active set sizing is uniform; (3) There is no steady state; parameters including the peers’ arrival rate, seeds’ departure
(4) Swarm performance is limited by upload bandwidth. The rate, connection probability and transmission bandwidth on the
authors built the model based on the expressions for the performance measures explicitly. They also carried out sensi-
probability of TFT reciprocation, expected downloading rate, tivity analysis of the performance metrics with respect to the
and expected upload rate. In the model, they considered two different parameters. They used a discrete-event simulator to
definitions of altruism. It can be defined as the difference validate the results obtained from analysis. They then extended
between expected upload rate and the download rate. Another the model to account for peers situated behind firewalls and
definition is any upload contribution that can be withdrawn discussed the impact on the system performance. Their results
without loss in download performance. Their result shows that clearly show that those peers not behind firewalls play an
the TFT policy is not robust, and the altruistic contribution important role in determining the overall system performance.
plays an important role in BitTorrent’s performance. They also discussed the effect of different chunk selection
The first heterogeneous fluid model of BitTorrent-like file algorithms at the peers on the file availability within the
sharing system was proposed by Lo Piccolo et al. in [43]. system. Through mathematical analysis they showed that the
To assess the effect of different access link capacities on rarest first policy indeed is the best to increase file availability.
BitTorrent-like file sharing system, they developed a fluid In practice this is true for low connection probability among
model with access links of two different capacity classes peers. However when ρ is large, this may cause a reduction in
by extending the fluid-flow model presented in [3]. To keep file availability. To improve file availability, they proposed a
the heterogeneity simple, they only considered links of two file availability enhancement algorithm which randomizes the
different link rates, viz. a high speed downloader and a low selection of the chunk to download, but still giving preference
speed downloader. Furthermore, they assumed that the links to the rarest chunk with higher probability. They showed
are symmetric, i.e., the upstream rate is the same as the that this enhancement improves the file availability in all
downstream rate. They analyzed the effects of bandwidth circumstances.
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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 like the connection probability ρ in [45]. References [25]
networks using the ‘uplink sharing’ approach, which is a and [24] studied the BitTorrent topology using measurement-
variation of the classic broadcasting problem. In their model, based studied, while [31] and [32] studied the same topic by
nodes have different upload capacities. They assumed that simulation. They observed the overlay structure and evaluated
the available upload capacity is shared equally amongst the the parameters which can impact the overlay topology. In
connections, and uploads are not interrupted until complete, addition, BitTorrent’s neighbor selection and the impact of
which means the rate is always positive. They showed that BitTorrent on ISP traffic were also measured by [8] and [26].
a simple and natural random strategy adopted by BitTorrent They show that the random neighbor selection and BitTorrent’s
results in a distribution time which is proportional to the locality-unawareness have a significant impact on the overall
optimal time achievable. performance.
Liao et al. [47] considered heterogeneity in a BitTorrent
system by distinguishing users into two classes, viz. high IV. I MPROVEMENTS TO B IT T ORRENT M ECHANISMS
bandwidth users who have high upload-link capacity, and The phenomenal success of BitTorrent has in its wake
low bandwidth users with low upload-link capacity. They attracted the attention of several researchers who examined its
proposed a mathematical model to predict the average file performance in detail as highlighted in the previous section.
download delay for the two classes of clients. This model Several researchers subsequently suggested improvements to
represents the system behavior in steady-state by matching existing BitTorrent mechanisms, or suggested new mecha-
the amount of data uploaded and downloaded by the peers. nisms to replace existing BitTorrent mechanisms with the aim
In addition, they designed a token-based TFT scheme. In of further improving its performance. In general, the suggested
this scheme tokens are used as a means for deciding how improvements can be viewed as belonging to two categories:
a peer uploads data to its neighbor. Tokens are earned by a (a) mechanisms that suggest improvements to BitTorrent’s
node by uploading to its neighbor. A node keeps track of overlay topology formation and maintenance, and (b) mod-
the number of tokens for its neighbors and makes its upload ifications to the piece exchange mechanisms in BitTorrent. In
decisions based on the tokens. By using a biased means for this section, we present an overview of the research literature
a low-bandwidth peer to earn more for uploading to a high- suggesting improvements to BitTorrent. We then categorize
bandwidth peer, the authors design a system which ensures these mechanisms as stated above.
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 A. Improvements to Overlay Topology
simulation results for several experiments conducted using the First, we review several proposals that aim to improve
BTSim simulator provided by [34] for different scenarios. The the performance of BitTorrent by changing the way peers
results show that the proposed mathematical model is quite establish their connections to other peers while joining, and
accurate in predicting the file download performance. 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
D. Summary and Comparison of Performance Studies
techniques are based on using the location of the peers to
In this section, we examine the performance studies from optimize the overlay topology. We classify the techniques into
the perspective of the different aspects of the BitTorrent two categories: (1) proximity aware techniques, and (2) ISP-
protocol that were studied. We summarize our findings in Ta- friendly techniques.
ble V, in particular highlighting how the different aspects were 1) Proximity Awareness: BitTorrent builds its overlay net-
studied using the different performance evaluation methods, work by selecting neighbors randomly. The randomized se-
with more or less the same conclusions. lection results in a significant divergence between the log-
All the three performance evaluation methods have focused ical overlay network and the underlying physical network
on evaluating the piece exchange mechanisms in BitTorrent. infrastructure [49]. This causes wastage of network resources
The general findings from the different approaches are in and potentially results in suboptimal downloading time. For
agreement with each other. In particular, it is found that the instance, two neighboring peers in the overlay may be in
TFT mechanism is very useful in ensuring co-operation among different countries far away from each other. In addition,
the peers. Furthermore, the general observation is that the selecting neighbors randomly also increases the probability
presence of altruistic peers encourages free-riding. Similarly, of having capacity-limited peers connected to high-capacity
the rarest first policy has been found by all methods to be very peers. To address these problems, some researchers have
effective in increasing download performance and decreasing explored the method of using proximity in the topology
the file download time. construction and maintenance in BitTorrent. The major aims
Topology related studies are feasible only through measure- of exploiting proximity are achieving an efficient usage of
ment and simulation, since these methods explicitly reflect the network resources, and reducing the individual downloading
effect of the topology on performance. In particular, the topol- time for the peers.
ogy’s effect is direct in measurement. In simulation studies, the Qureshi [49] was among the first to suggest the use of
topology is usually abstracted at the overlay topology level. proximity in the BitTorrent overlay network. He pointed out
Analytical modeling does not explicitly track the topology, that the present overlay construction algorithms ignore the
but its influence may be reflected through system parameters problem of matching the overlay network and the underlying
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 152 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010

TABLE V
S UMMARY OF M ETHODS U SED IN S TUDYING B IT T ORRENT P ERFORMANCE

Issues Measurement Simulation Analysis

TFT(Incentive The TFT policy works well [18]. The TFT policy fails to prevent The incentive mechanism are
Piece Exchange

mechanism) The replacement with a bit level unfairness [29]. efficient [3]. Optimistic unchoking
TFT solution is not appropriate may encourage free-riders [3].
[22]. The TFT mechanism in- BT’s mechanisms can guard
creases the cooperative behavior against free-riding successfully
and discourages free-riding, but if [39]. TFT policy is not the major
there are a large number of seeds, it cause for good performance; it
may not work effectively to prevent is the altruistic contribution by
free-riding [20]. high-capacity peers [42]
Piece Selection The rarest first policy works well Rarest first piece selection strategy Rarest first strategy is still the best
[18]. The rarest first policy pre- gives balanced performance [28]. way to increase file availability
vents the reappearance of rare The rarest first policy is critical to [45].
pieces and avoids the last pieces ensure the new joined peers have
problem [22]. something to exchange with others
[29].
Topology Choking algorithm can cluster The overlay is not a random graph
Overlay Topology

peers with similar bandwidth [24]. [31] The overlay topology can be
There is no evidence of clustering affected by the choice of peer set
except the initial stage [25]. size and the percentage of outgoing
connections [32].

Neighbor The random neighbor selection in The peer selection strategy may
Selection BT causes low transmission rate lead to the mismatched peering and
and high operating cost [26]. BT is make BT deviate more from the
locality-unaware, which increases optimum [33].
the ISPs’ cost [8].
Impact of Fire- Peers behind a firewall have low A large number of NATed peers de- Presence of non-firewalled peers
wall/NAT bandwidth usage [23]. Peers be- crease the robustness of the overlay essential to good system perfor-
hind a firewall often disconnect to attacks significantly [31] mance [45].
early [23].

network. He suggested that peers that are close by in the real specified as a 3-tuple representing the three coordinates, which
world should be close by in the overlay network. Formally, could be x-coordinate, y-coordinate and z-coordinate. When a
for any two nodes A and B, he specified their round-trip node joins into the network, the coordinates will be assigned
communication latency as rtt[A,B] . Given three nodes A, B to it by the tracker. Then, the node will measure the Euclidean
and C, if rtt[A,B] < rtt[A,C] , then P (A, B) > P (A, C), distance between other nodes and itself, and proactively select
where P (A, B) is the probability that A and B are peers in the nodes with the shortest Euclidean distance as its neighbors.
overlay network. Several techniques including using network Their results show that using location proximity in BitTorrent
coordinates coupled with a probabilistic flooding algorithm can decrease the file downloading time about 11.3% for 1000
were exploited in the topology construction and maintenance. nodes torrent, and improve the network utilization, especially
This was done in two phases: (1) estimating the node’s for moderate size torrent.
network coordinates; (2) using a gossip-like protocol for near Koo et al. [52] proposed a neighbor selection strategy
neighbor discovery. The expected result was that an overlay for hybrid P2P network like BitTorrent that favors picking
network that closely matches the underlying network topology neighbors with the most mutually disjoint content. In [52],
in terms of node proximity should yield shorter downloading they modeled peers as an undirected graph. In addition,
time and more efficient usage of network resources. they presented a genetic-algorithm-based method for neigh-
Zhang et al. [50] analyzed location proximity in BitTorrent bor selection. They use ns-2 [53] package to simulate the
system not only based on topology proximity, but also for system. In their simulation model, there are three classes of
cooperative proximity. They indicated that both neighbor peers with different bandwidths. They compared the proposed
selection and piece selection phases are proximity-unaware. neighbor selection strategy with the existing strategy using two
This can cause problems such as topology mismatch, and long metrics: system throughput and average downloading time.
average downloading time. They considered a proximity aware Their results show that the new neighbor selection algorithm
topology construction algorithm. Their approach consists of can improve system performance by increasing the content
two steps. In the first step, they give each node a synthetic availability for peers from their neighbors.
network coordinate by using the algorithms in [51]. In the The proposed methods of neighbor selection by proximity
second step, each node selects its neighbors by their Euclidean [49], [50] are still oblivious to ISP boundaries. Peers selected
distance. Their simulations are conducted using a modified with proximity in mind may belong to different ISPs, thus still
version of BTSim [34] extended to exploit proximity in over- not addressing the cross-ISP traffic problem.
lay topology construction. They built a 3-dimensional space 2) ISP-Friendliness: The wide use of BitTorrent has put
representation of the network, where each node’s location was ISPs in a dilemma [54]! On the one hand, BitTorrent, with
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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 authors to improve the piece exchange mechanisms in BitTor-
revenue for ISPs. On the other hand, BitTorrent generates a rent to further improve download performance, punish free-
large amount of cross-ISP traffic and makes the ISPs’ cost riders and address unfairness. The first set of mechanisms aims
increase significantly since BitTorrent’s implementations ig- to promote collaboration among peers using the underutilized
nore the underlying Internet topology and ISP costs. Lei et al. download bandwidth of peers to help others. The second set
[26] mentioned that BitTorrent’s fast downloading experience of mechanisms address the issue of free-riding through several
comes at a large cost to the network in terms of resources. This approach to identify and punish free-riders. The third set of
may consequently affect other Internet services supported by mechanisms address the issue of unfairness in BitTorrent by
the ISPs. suggesting improvements to the unchoke mechanisms.
To address the cross-ISP problem, Bindal et al. [9] designed 1) Collaboration among Peers: BitTorrent’s piece ex-
a new algorithm based on biased neighbor selection, in which change mechanism is designed to enforce fairness among
a peer chooses its neighbors mostly from peers within the peers through the TFT mechanism. Garbacki et al. [56]
same ISP, and only selects a few from the outside its ISP pointed out three limitations of TFT: (1) the newcomers are
domain. This is because peers in the same ISP are highly bootstrapped at the bandwidth cost of the existing peers; (2)
connected. However, some connectivity to peers outside the no incentives for seeding; (3) peers with asymmetric Internet
ISP boundaries needs to be maintained in order to obtain new connections cannot fully use their downloading links since
blocks. They indicated that the biased neighbor selection can they are forced to download at the speed of their uploading
be implemented either by changing the tracker and client or by link.
employing P2P traffic shaping devices. When there is a request Garbacki et al. [57] proposed a protocol named 2Fast which
from a peer for neighbor list, the tracker will choose k external extended the bartering model of BitTorrent. In 2Fast, a peer
peers and 35 − k internal peers to feed back to the requesting with idle bandwidth can join a download in progress as a
peer. If there are less than 35 − k internal peers, the tracker helper to assist a peer (hereafter named the collector) and fetch
will ask the requesting peer to contact it again after certain missing fragments of the file being downloaded and thereafter
duration. The tracker can know the peers’ ISP location either upload them to the collector. A collector recruits peers that are
by using the AS mapping or IP address to identify the ISP, willing to act as its helpers with the promise that the bandwidth
or add a new header which contains the locality tag to HTTP contributed by the helper will be returned in the future. The
proxy. Biased neighbor selection can also be implemented by authors used social incentives to enforce the delivery on the
P2P traffic shaping devices, which are situated alongside the promise to return bandwidth in the 2Fast system. In [58], the
edge routers of the ISPs and use deep packet inspection to same authors exploited social phenomena like friendships to
identify P2P traffic. To understand the process of neighbor enable collaboration. Here a collector finds helpers just like
selection, the authors designed a discrete-event simulator and people collaborate with friends in communities. Through rig-
evaluated it over a network configuration with 14 ISPs. They orous analysis, the authors show that adding a new helper will
considered the flash crowd phase. The main metrics evaluated not help if the collector’s downloading bandwidth is already
include the download time, and ISP traffic redundancy which saturated. They computed the minimum number of helpers
is the average number of times the blocks of the downloading that fills the collector’s bandwidth. Finally, they conducted
file travels into the ISP until all peers inside the ISP finish their experiments to compare the BitTorrent protocol enhanced with
download. The simulation result shows that biased neighbor the 2Fast protocol against the standard BitTorrent protocols
selection can reduce the cross-ISP traffic and maintain the using several metrics such as the downloading speedup, which
nearly optimal performance of BitTorrent. is the ratio between the downloading time of a peer acting on
Yamazaki et al. [55] proposed a series of strategies to reduce its own and the downloading time of a collector supported by
ISP costs called Cost-Aware BitTorrent (CAT). In CAT, the its helpers. The results show that the use of 2Fast protocol
peers first acquire ISP cost information. Thereafter, path costs improves the download speed by up to a factor of 3.5 in
between any two peers can be computed using the ISP costs. comparison to the standard BitTorrent without helpers. Fig. 3
The tracker selects peers in the list of peers it returns based depicts the collaboration between a collector and its helpers
on the shortest cost to the requesting peer. The inter-peer in the 2Fast protocol.
cost is used by a peer in the selection of candidate peers Subsequently, Garbacki et al. [56] extended the 2Fast pro-
from which to download and upload pieces. The strategy is to tocol and proposed a novel mechanism in which incentives
minimize the cost incurred in the transactions, consequently are built around bandwidth rather than content. Bandwidth
minimizing the ISP costs. Through simulations for different is unrelated to the interests of peers, so it is more suitable
scenarios they show that CAT can reduce the ISP cost and to be a trading unit. In the bandwidth-exchange incentive
improve the performance. 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
B. Improvements to Piece Exchange Mechanism this protocol peers choose the bandwidth borrowers with the
As already mentioned in Sec. II, the tit-for-tat (TFT) mech- highest chance of returning the borrowed bandwidth. ATFT
anism ensures that peers contribute as much as they download. employs the exploration and selection operations to select
Similarly, optimistic unchoking (OU) enables new peers to get peers with the highest contributions as the borrowers. Explo-
started, and enables peers to discover better matched peers for ration extends the borrowers set by opening opportunities for
cooperation. In this section, we review several suggestions by bandwidth exchange, and selection reduces the borrowers set
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 154 IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. 12, NO. 2, SECOND QUARTER 2010

Non-contributor peer

Requesting (busy)
contributor

Idle contributor acting as

helper

Collector

Contribution from helper

Tit-for-tat

Collaboration
Collaborative Download Non-collaborative Download
Contributors set
Collector Dowload Completed

Helper Downloading Fig. 4. The Collector and helpers in ATFT protocol

Fig. 3. Collector and helpers in 2Fast system

when they are downloading. They also implemented a simu-
lator for the original BitTorrent and BitTorrent with helpers,
by removing the least promising peers. Fig. 4 illustrates the using which they compared the average downloading time.
ATFT protocol from the view of the collector. The authors They presented several conclusions: (1) helpers are almost as
pointed out four benefits of ATFT. Firstly, newcomers pay the effective as seeds though they only download a tiny fraction
cost of bootstrapping by themselves. Secondly, ATFT provides of the file; (2) helpers that download too many pieces will hurt
an effective incentive for content seeding. Thirdly, peers can the system performance; (3) average downloading time of a
fill their downlinks completely with the helpers’ bandwidth. BitTorrent system with helpers is much shorter than that of
Finally, it can be used to provide a certain level of anonymity. the BitTorrent system without helpers; (4) it is more effective
They conducted several simulation experiments using ATFT. for BitTorrent to have a lot of helpers than to have a few extra
In their simulations, a peer maintains a single value for seeds.
keeping track of the bandwidth obtained from contributors to 2) Addressing Free-Riding: A free-rider is a node that
account for bandwidth contributions. The results show that downloads pieces from other peers but does not upload any
ATFT improves the average download bandwidth of a peer by pieces to others. Conventional P2P systems lack incentive
a factor of 2 to 6. mechanisms and are vulnerable to free-riding. Incentive mech-
The above protocols do not explicitly specify how many anisms encourage cooperation among peers by rewarding
pieces the helpers should download to make the system work contributors and punishing free-riders. Although BitTorrent
well. On one hand, if helpers download too many pieces, it has incentive mechanisms such as TFT, many researchers
would incur excessive transmission cost; on the other hand, pointed out that they are not effective in preventing free-riding.
if the helper downloads too few pieces, it may not be very Jun et al. [60] evaluated the original incentive mecha-
useful. Wang et al. [59] showed that helpers can work as nism of BitTorrent using measurements on a P2P network
effectively as seeders if they download only a tiny fraction set up on PlanetLab [61]. They view the peer behavior in
of the file. Helpers in [59] are resource-rich nodes with spare BitTorrent as akin to the iterated prisoner’s dilemma. They
uploading bandwidth that can be used to increase the total found that free-riders are not effectively punished, and peers
system uploading bandwidth and hence easing the bottleneck. that contribute to others are not rewarded appropriately. To
They proposed an efficient strategy for utilizing the spare address this problem, they proposed a new mechanism which
uploading capacity, where a helper joins a swarm just like is more robust against free-riders. In their new mechanism,
a regular peer, and only downloads a small number, k pieces, they defined peers’ upload amount u and download amount
of the file. Once a helper downloads k pieces, it will stop d for each link. Then, they define a concept named deficit,
downloading, and at this time becomes a microseed. The which is u−d. A peer ensures that the deficit is always upper-
microseed contacts the tracker and gets a list of peers who bounded by the formula: u − d ≤ f.c, where the constant c
need help. The microseed implements choking and unchoking is the size of a fragment, and f (≥ 1) is called a nice factor.
algorithms. Once the microseed knows that a neighbor has The factor determines the amount that a peer is willing to risk
already downloaded all the pieces that it possesses, it will for a chance to establish cooperation. Peers upload evenly
disconnect from the neighbor. Helpers download only a small to all links as much as they can under this condition. The
number of pieces because they are not interested in the file. authors evaluated the new mechanism experimentally through
Likewise, downloading a large part of the file is a waste of measurements using PlanetLab. The experiments consisted of
helpers’ bandwidth. The authors used a fluid model based on one tracker, one seed and 170 leechers running BitTorrent
the model presented in [3] to analyze the system performance. 4.0.0. All of the leechers begin to download at the same time.
They made several assumptions: the peer arrival process is During the experiment, the authors evaluated two metrics:
Poisson; peers are homogeneous with the same uploading the downloading time and the uploading amount. The results
and downloading bandwidth; peers download at the same rate show that the original incentive mechanism in BitTorrent is
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 155

suboptimal and the proposed mechanism is more robust at and only if Uab ≤ Dab + Δ , where Uab and Dab are the
preventing free-riding. uploaded and the downloaded blocks from node A to/from
Sirivianos et al. [62] presented a new free-riding technique node B respectively, and Δ is the unfairness threshold for
named the large view exploit. In this approach they modified this connection. This ensures that the maximum number of
the BitTorrent client to: (a) not upload content to others, (b) extra blocks served by a node is bounded by dΔ where d is
periodically contact the tracker behaving as a new peer and the size of its neighborhood. The authors used their simulator
obtain a new list of peers, and (c) establish connections with to evaluate the performance of the new techniques based
peers in the new list, thus exploiting their optimistic unchoke on three metrics: mean uploading utilization, unfairness, and
contributions. The authors performed several experiments both mean downloading time. The results showed that the proposed
with PlanetLab resident torrents and public torrents to study methods, QBE and pairwise block-level TFT can effectively
the effectiveness of the large view exploit. Their results show address the unfairness towards high-capacity nodes, and also
that the modified peer can easily achieve better performance improve the upload link utilization. We should note that QBE
than the traditional peers. The authors then suggest a modi- is somewhat idealistic since the reliable bandwidth estimation
fication to the BitTorrent tracker and clients to address this is far from a trivial exercise [64]. The block-level TFT may
problem. The aim is to give the clients and the tracker a result in a reduction of uploading utilization because peers
consistent view of the whole swarm. The approach is to use can potentially cease to upload when the block-level TFT
a peer’s IP address together with a pseudo-random number constraint is not satisfied.
sequence to enable other peers to attempt to identify free- Thommes et al. [64] simulated a simplified version of
riders. The modification requires the tracker to perform a BitTorrent protocol using Matlab in order to understand the
larger amount of housekeeping, and may impact the scalability fairness properties. Based on their evaluation, they proposed
of the approach. The authors are continuing to investigate this three methods to improve BitTorrent fairness. The three new
approach further to improve scalability. mechanisms are as follows: The first method is conditional
Chow et al. [63] presented a novel approach from the optimistic unchoke. In this method a peer performs an op-
perspective of using the seed capacity appropriately with the timistic unchoke only when it’s IFR is larger than 1. The
goal of reducing the free-riders. Several of the measurements second method is multiple connection chokes, which allows
and simulation studies of BitTorrent reveal that the leechers’ peers to choke/unchoke not just one connection, but multiple
downloading rate is slow at the beginning when they have connections in each round. A peer computes the Connection
few chunks to exchange with others. Similarly, it can be Fairness, which is the ratio of the peer’s uploading rate to
slow at the end due to the inability of the peer in finding a specific connected peer to the downloading rate from that
neighbors possessing the few missing chunks that it requires peer, for each of the five peers to which it is connected. They
to complete the download. So the authors proposed a simple evaluate two parameters: (a) Threshold Ratio, which is the
method to prioritize the use of seeding capacity to certain largest value the Connection Fairness can assume before the
portions of the file downloading process. They used two corresponding upload is choked; (b) Maximum Chokes (MC),
ways to choose neighbors to unchoke: (1) Sort-based: a seed which is the largest number of uploads a peer can choke
sorts its neighbors according to the number of chunks each in each round. If the number of connection fairness values
possesses. Then it unchokes N of them based on the sorting exceeding the Threshold Ratio is larger than the Maximum
order; (2) Threshold-based: a seed unchokes N neighbors Chokes, the peer will choke the connections which are unfair.
with [0..K/2]% or [(100 − K/2)..100]% of the chunks. Their Otherwise, it will choke only MC of the peers. The third mech-
experiments use BTSim [9] with the following modifications: anism is variable number of outgoing connections, in which
(1) nodes could be seeds with different seeding approach; (2) the number of connections a peer maintains is variable, instead
nodes act as free-riders; (3) continuous node arrivals. Their of a fixed number. The number of outgoing connections varies
experimental results show that the new seed capacity prioriti- depending on the peer’s upload capacity. They also allow the
zation approach not only discourages free-riding behavior, but peer’s upload capacity dedicated to each connection to be
also improves the performance of contributing leechers. variable. Each peer evaluates its set of outgoing and incoming
3) Improving Fairness: Bharambe et al. [29] found that connections every 10 seconds. In each iteration, the peer
BitTorrent’s optimistic unchoke mechanism can result in sys- constructs a list Lnd of peers to which it is currently uploading,
tematic unfairness where a high-capacity node might end but from which the peer is not receiving any data. The peer
up uploading a larger amount of data than it downloads. will choke the peers in list Lnd . Then, the peer constructs
They proposed two strategies to reduce this unfairness. The another list Lnu of peers from which it is downloading, but to
first one named Quick bandwidth estimation (QBE) tries to which it is not uploading. When |Lnu | is larger than |Lnd |, it
circumvent the need for optimistic unchokes. This is feasible starts uploading to a random set of |Lnd | peers in Lnu . When
if the nodes can quickly estimate the uploading bandwidths |Lnu | is smaller than |Lnd |, it begins to upload to all peers
for all its neighboring peers. QBE can estimate a neighbor’s in Lnu , and randomly choose |Lnd | − |Lnu| peers to unchoke
upload bandwidth based on the history of past interactions optimistically.
with it. Alternately, the packet-pair principle can be used to They defined the Instantaneous Fairness Ratio (IFR) for
estimate the bandwidth. The second proposed algorithm is each peer as the ratio of data uploaded to data downloaded
pairwise block-level TFT, which focuses on fairness of the during the previous 10 seconds. If a peer’s IFR is less than 1,
number of blocks transferred instead of uploading rates. Node it means the peer is downloading an excessive amount. Oth-
A allows its neighbor B to download a block from it if erwise, it tells us that the peer is downloading an insufficient
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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. [10] C. Gkantsidis and P. Rodriguez, “Network coding for large scale content
The results show that the three modifications provide some distribution,” in IEEE INFOCOM. Miami, USA: IEEE Computer
Society, March 2005, pp. 2235–2245.
level of improvement in BitTorrent’s fairness. [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.
C. Summary
[12] R. Sherwood, R. Braud, and B. Bhattacharjee, “Slurpie: a cooperative
Table VI summarizes the improvements to BitTorrent’s bulk data transfer protocol,” INFOCOM 2004. Twenty-third AnnualJoint
mechanisms that are presented earlier in this section. As can Conference of the IEEE Computer and Communications Societies,
vol. 2, pp. 941–951, March 2004.
be seen from the table, each suggested improvement addresses [13] E. K. Lua, J. Crowcroft, M. Pias, R. Sharma, and S. Lim, “A survey and
one of the issues with BitTorrent mechanisms. No studies of comparison of peer-to-peer overlay network schemes,” IEEE Commun.
how these improvements interact with each other if included Surveys Tutorials, vol. 7, no. 2, pp. 72–93, Quarter 2005.
[14] S. Androutsellis-Theotokis and D. Spinellis, “A survey of peer-to-peer
together in BitTorrent are available. It would be interesting to content distribution technologies,” ACM Comput. Surv., vol. 36, no. 4,
explore the combined effects of the different improvements pp. 335–371, 2004.
suggested to see if they indeed work collaboratively to further [15] M. Hosseini, D. Ahmed, S. Shirmohammadi, and N. Georganas, “A
survey of application-layer multicast protocols,” IEEE Commun. Surveys
improve performance, or impede each other. Tutorials, vol. 9, no. 3, pp. 58–74, Quarter 2007.
[16] D. Chopra, H. Schulzrinne, E. Marocco, and E. Ivov, “Peer-to-peer
V. C ONCLUSION overlays for real-time communication: Security issues and solutions,”
IEEE Commun. Surveys Tutorials, vol. 11, no. 1, pp. 4–12, 2009.
This paper presented a survey of BitTorrent performance. [17] R. Ranjan, A. Harwood, and R. Buyya, “Peer-to-peer-based resource
First we reviewed the performance studies of the original discovery in global grids: A tutorial,” IEEE Commun. Surveys Tutorials,
vol. 10, no. 2, pp. 6–33, 2008.
BitTorrent’s protocols including measurement, simulation, and [18] M. Izal, G. Urvoy-keller, E. W. Biersack, P. A. Felber, A. A. Hamra,
analytical modeling. We then summarized the findings of and L. Garces-Erice, “Dissecting BitTorrent: Five months in a torrents
these studies. Next, we summarized some of the suggested lifetime,” in Passive and Active Measurements (PAM), LNCS, vol. 3014,
April 2004, pp. 1–11.
improvements to BitTorrent’s mechanisms in order to further [19] A. Bellissimo, B. N. Levine, and P. Shenoy, “Exploring the use of
improve its performance. BitTorrent as the basis for a large trace repository,” Technical report,
We note that the innovative mechanisms introduced by Bit- University of Massachusetts Amherst, Dept. of Computer Science, Tech.
Rep., June 2004.
Torrent has influenced not only P2P file sharing applications, [20] N. Andrade, M. Mowbray, A. Lima, G. Wagner, and M. Ripeanu,
but also served as the foundation for several other applications “Influences on cooperation in BitTorrent communities,” in P2PECON
including P2P streaming and P2P Voice over IP. For example, ’05: Proceedings of the 2005 ACM SIGCOMM workshop on Economics
of peer-to-peer systems. New York, NY, USA: ACM, 2005, pp. 111–
BitTorrent based streaming applications are described in [65], 115.
[66]. Thus we see the study of BitTorrent and its performance [21] A. Rasti and R. Rejaie, “Understanding peer-level performance in
as a very useful step not only to understand its performance, BitTorrent: A measurement study,” Computer Communications and
Networks, 2007. ICCCN 2007. Proceedings of 16th International Con-
but also as a means of learning from the experience to enable ference on, pp. 109–114, Aug. 2007.
better design of similar protocols for other applications. [22] A. Legout, G. Urvoy-Keller, and P. Michiardi, “Rarest first and choke
algorithms are enough,” in IMC ’06: Proc. 6th ACM SIGCOMM
ACKNOWLEDGMENT conference on Internet measurement. New York, NY, USA: ACM,
2006, pp. 203–216.
The work described in this paper has been supported by [23] K. andré Skevik, V. Goebel, and T. Plagemann, “Analysis of BitTorrent
HK RGC under RGC-Competitive Earmarked Research Grant and its use for the design of a p2p based streaming protocol for a hybrid
cdn,” Delft University of Technology, Tech. Rep., 2004.
HKUST 617907. The authors would like to thank all the [24] A. Legout, N. Liogkas, E. Kohler, and L. Zhang, “Clustering and sharing
anonymous reviewers for their invaluable suggestions. 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,
R EFERENCES pp. 301–312.
[1] “Bittorrent,” http://www.bittorrent.com/. [25] C. Dale, J. Liu, J. Peters, and B. Li, “Evolution and enhancement of
[2] B. Cohen, “Incentives build robustness in BitTorrent,” in First Workshop BitTorrent network topologies,” Quality of Service, 2008. IWQoS 2008.
on Economics of Peer-to-peer Systems, Berkeley, USA, June 2003. 16th International Workshop on, pp. 1–10, June 2008.
[3] D. Qiu and R. Srikant, “Modeling and performance analysis of [26] Y. Lei, L. Yang, Q. Jiang, and C. Wu, “Experimental views on neighbor
BitTorrent-like peer-to-peer networks,” in SIGCOMM ’04: Proc. 2004 selection in BitTorrent,” Network and Parallel Computing Workshops,
conference on Applications, technologies, architectures, and protocols 2007. NPC Workshops. IFIP International Conference on, pp. 813–818,
for computer communications. New York, NY, USA: ACM, 2004, pp. Sept. 2007.
367–378. [27] “Bittornado website,” http://www.bittorrent.com/, 2007.
[4] “Gnutella,” http://www.gnutelliums.com/. [28] P. A. Felber and E. W. Biersack, “Self-scaling networks for content dis-
[5] “Kazaa,” http://www.kazaa.com/. tribution,” in Int. Workshop on Self-* Properties in Complex Information
[6] J. A. Pouwelse, P. Garbacki, D. H. J. Epema, and H. J. Sips, “The Systems, Bertinoro, Italy, May-June 2004.
BitTorrent p2p file-sharing system: Measurements and analysis,” in [29] A. R. Bharambe, C. Herley, and V. N. Padmanabhan, “Analyzing and
IPTPS’05, 2005, pp. 205–216. improving a BitTorrent networks performance mechanisms,” INFOCOM
[7] L. Ellis, “BitTorrent’s swarms have a deadly bite on broadband nets,” 2006. 25th IEEE International Conference on Computer Communica-
Multichannel news, 2006. tions. Proceedings, pp. 1–12, April 2006.
[8] T. Karagiannis, P. Rodriguez, and K. Papagiannaki, “Should internet ser- [30] E. Adar, “Drawing crowds and bit welfare,” SIGecom Exch., vol. 5,
vice providers fear peer-assisted content distribution?” in IMC ’05: Proc. no. 4, pp. 31–40, 2005.
5th ACM SIGCOMM conference on Internet Measurement. Berkeley, [31] A. Al Hamra, A. Legout, and C. Barakat, “Understanding the properties
CA, USA: USENIX Association, 2005, pp. 6–6. of the BitTorrent overlay,” INRIA, Sophia Antipolis, Tech. Rep., July
[9] R. Bindal, P. Cao, W. Chan, J. Medved, G. Suwala, T. Bates, and 2007.
A. Zhang, “Improving traffic locality in BitTorrent via biased neighbor [32] G. Urvoy-Keller and P. Michiardi, “Impact of inner parameters and
selection,” in ICDCS ’06: Proc. 26th IEEE International Conference overlay structure on the performance of BitTorrent,” INFOCOM 2006.
on Distributed Computing Systems. Washington, DC, USA: IEEE 25th IEEE International Conference on Computer Communications.
Computer Society, 2006, p. 66. Proceedings, pp. 1–6, April 2006.
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.
 XIA and MUPPALA: A SURVEY OF BITTORRENT PERFORMANCE 157

TABLE VI
S UMMARY OF I MPROVED M ECHANISMS S UGGESTED FOR B IT T ORRENT

Mechanisms Suggested Improvements Features of Improvement Result Ref.

Random Neighbor selection by Synthetic network coordinates, Eu- Decrease the file downloading time [49],
Overlay Topology

Neighbor proximity clidean distance and increase network utilization [50]

Selection

Neighbor selection by con- Select neighbor by the most mutually Maximize content availability from [52]
tent disjoint content neighbors
ISP- Biased neighbor selection Choose most neighbors inside the ISP Reduce the cross-ISP traffic [9]
unawareness
Cost-aware Selection Select neighbors according to the cost Reduce the ISP cost [55]
between peers
No Collabo- Collaborative Download Helpers: peers with idle bandwidth can Improve download speed; better re- [57],
Piece Exchange

rative Mech- help other peers downloading source utilization [58],

anism [59]

ATFT Incentives are built on bandwidth in- Several benefits including for new [56]
stead of content peers, seeds; better resource uti-
lization
Helpers Helpers download a small fraction of Improved download performance [59]
file and act as microseeds for peers
Addressing New incentive mechanism Peers upload evenly to all links as The new proposed mechanism is [60]
Free-Riding much as they can under the condition more robust to prevent free-riding
u − d <= f.c.
Addressing the problems Using peer’s IP address to track free- Better ability to identify free-riders [62]
of large view exploit riding behavior
Seed capacity utilization Use the seeding capacity to certain por- Discourage free-riders’ behavior, [63]
tions of a file downloading process Improve the leechers’ Performance
Improving QBE; Pairwise block-level QBE: uploading bandwidths can be es- Address the problems of low up- [29]
Fairness TFT timated quickly; Pairwise block-level link utilization and unfairness
TFT: the maximum number of extra
blocks served by a node is bounded
Conditional optimistic un- Conditional optimistic unchoke: only Improve fairness [64]
choke; Multiple connec- when a peer’s IFR is larger than 1,
tion chokes; Variable num- it performs an optimistic unchoke al-
ber of outgoing connec- gorithm; Multiple connection chokes:
tions peers can choke/unchoke multiple con-
nections in each round; Variable number
of outgoing connections: the number of
connections a peer maintains is variable

[33] G. Wu and T. cker Chiueh, “How efficient is BitTorrent?” in Multimedia like peer-to-peer networks,” Services Computing, 2006. APSCC ’06.
Computing and Networking 2006, S. Chandra and C. Griwodz, Eds., IEEE Asia-Pacific Conference on, pp. 534–538, Dec. 2006.
vol. 6071, no. 1. SPIE, 2006, p. 60710O. [Online]. Available: [40] Y. Tian, D. Wu, and K. W. Ng, “Modeling, analysis and improvement
http://link.aip.org/link/?PSI/6071/60710O/1 for BitTorrent-like file sharing networks,” INFOCOM 2006. 25th IEEE
[34] A. Bharambe, C. Herley, and V. Padmanabhan, “Mi- International Conference on Computer Communications. Proceedings,
crosoft research simulator for the BitTorrent protocol,” pp. 1–11, April 2006.
http://research.microsoft.com/projects/btsim/, 2006. [41] P. Saraswat and P. Batra, “An empirical performance evaluation and
[35] X. Yang and G. de Veciana, “Performance of peer-to- modelling of BitTorrent peer-to-peer file sharing system using queuing
peer networks: Service capacity and role of resource sharing network models,” Research project, Advanced Learning and Research
policies,” Performance Evaluation, vol. 63, no. 3, pp. 175 Institute, Switzerland, 2007.
– 194, 2006, p2P Computing Systems. [Online]. Avail- [42] M. Piatek, T. Isdal, T. Anderson, A. Krishnamurthy, and A. Venkatara-
able: http://www.sciencedirect.com/science/article/B6V13-4FJGW2F- mani, “Do incentives build robustness in BitTorrent?” in NSDI’07,
1/2/be7c36ff2970dd2cea798f2de6200a40 Cambridge, MA, April 2007, pp. 1–14.
[36] L. Guo, S. Chen, Z. Xiao, E. Tan, X. Ding, and X. Zhang, “A [43] F. Lo Piccolo and G. Neglia, “The effect of heterogeneous link capacities
performance study of BitTorrent-like peer-to-peer systems,” IEEE J. Sel. in BitTorrent-like file sharing systems,” Peer-to-Peer Systems, 2004.
Areas Commun., vol. 25, no. 1, pp. 155–169, Jan. 2007. International Workshop on Hot Topics in, pp. 40–47, Oct. 2004.
[37] D. Arthur and R. Panigraphy, “Analyzing the efficiency of BitTorrent [44] F. Clévenot-Perronnin, P. Nain, and K. W. Ross, “Multiclass
and related peer-to-peer networks,” in Proc. Seventeenth annual ACM- p2p networks: Static resource allocation for service differentiation
SIAM symposium on Discrete algorithms, January 2005, pp. 961–969. and bandwidth diversity,” Performance Evaluation, vol. 62,
[38] R. Kumar and K. Ross, “Peer-assisted file distribution: The minimum no. 1-4, pp. 32 – 49, 2005, performance 2005. [On-
distribution time,” Hot Topics in Web Systems and Technologies, 2006. line]. Available: http://www.sciencedirect.com/science/article/B6V13-
HOTWEB ’06. 1st IEEE Workshop on, pp. 1–11, Nov. 2006. 4GV9S5V-5/2/182bf23d02fcbeff01e45a97ed891a9a
[39] J. Yu, M. Li, F. Hong, and G. Xue, “Free-riding analysis of BitTorrent- [45] B. Fan, D.-M. Chiu, and J. Lui, “Stochastic analysis and file availability
Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 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. [60] S. Jun and M. Ahamad, “Incentives in BitTorrent induce free riding,” in
IWQoS 2006. 14th IEEE International Workshop on, pp. 30–39, June P2PECON ’05: Proc. 2005 ACM SIGCOMM workshop on Economics of
2006. peer-to-peer systems. New York, NY, USA: ACM, 2005, pp. 116–121.
[46] J. Mundinger, R. Weber, and G. Weiss, “Optimal scheduling of peer- [61] “Planetlab,” http://www.planet-lab.org/.
to-peer file dissemination,” Journal of Scheduling, vol. 11, no. 2, pp. [62] M. Sirivianos, J. Park, R. Chen, and X. Yang, “Freeriding in BitTorrent
105–120, April 2008. networks with the large view exploit,” in IPTPS’07, 2007.
[47] W.-C. Liao, F. Papadopoulos, and K. Psounis, “Performance analysis [63] A. L. Chow, L. Golubchik, and V. Misra, “Improving BitTorrent: A
of BitTorrent-like systems with heterogeneous users,” Performance simple approach,” in Proceedings of IPTPS, 2008.
Evaluation, vol. 64, no. 9-12, pp. 876 – 891, 2007. [On- [64] R. Thommes and M. Coates, “BitTorrent fairness: analysis and improve-
line]. Available: http://www.sciencedirect.com/science/article/B6V13- ments,” in Workshop Internet, Telecom. and Signal Proc. (WITSP’05),
4P192CX-2/2/d13c7120e07234e69160c85b5b861f8d Dec 2005.
[48] M. Bertoli, G. Casale, and G. Serazzi, “An overview of the jmt queuing [65] S. Tewari and L. Kleinrock, “Analytical model for BitTorrent-based
network simulator,” Politecnico di Milano, Tech. Rep., 2007. live video streaming,” Consumer Communications and Networking
[49] A. Qureshi, “Exploring proximity based peer selection in a BitTorrent- Conference, 2007. CCNC 2007. 4th IEEE, pp. 976–980, Jan. 2007.
like protocol,” MIT 6.824 student project, 2004. [66] A. Vlavianos, M. Iliofotou, and M. Faloutsos, “BiToS: Enhancing
[50] L. Zhang, J. K. Muppala, and W. Tu, “Exploiting proximity in coop- BitTorrent for supporting streaming applications,” INFOCOM 2006.
erative download of large files in peer-to-peer networks,” International 25th IEEE International Conference on Computer Communications.
Conference on Internet and Web Applications and Services, (ICIW’07), Proceedings, pp. 1–6, April 2006.
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 Raymond Lei. Xia is a M. Phil. student in The Hong Kong University of
Computer Systems, vol. 22, no. 7, pp. 732 – 741, 2006. [On- Science and Technology, Clear Water Bay, Kowloon, Hong Kong.
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 Jogesh K. Muppala (M’85-SM’02) received the Ph. D. degree in Electrical
and p2p users cooperate for improved performance,” ACM SIGCOMM Engineering from Duke University, Durham, NC in 1991. He is currently an
Computer Communication Review, vol. 37, pp. 29–40, 2007. associate professor in the Department of Computer Science and Engineering,
[55] S. Yamazaki, H. Tode, and K. Murakami, “CAT: A cost-aware BitTor- The Hong Kong University of Science and Technology, Clear Water Bay,
rent,” in 32nd IEEE Conference on Local Computer Networks (LCN Kowloon, Hong Kong. He was previously a Member of the Technical Staff
2007), Oct 2007, pp. 226–227. at Software Productivity Consortium (Herndon, Virginia, USA) from 1991 to
[56] P. Garbacki, D. Epema, and M. van Steen, “An amortized tit-for-tat 1992, where he was involved in the development of modeling techniques
protocol for exchanging bandwidth instead of content in p2p networks,” for systems and software. While at Duke University, he participated in
Self-Adaptive and Self-Organizing Systems, 2007. SASO ’07. First the development of two modeling tools, the Stochastic Petri Net Package
International Conference on, pp. 119–128, July 2007. (SPNP) and the symbolic Hierarchical Automated Reliability and Performance
[57] P. Garbacki, A. Iosup, D. Epema, and M. van Steen, “2fast: Collaborative Evaluator (SHARPE), both of which are being used in several universities
downloads in p2p networks,” in P2P ’06: Proc. Sixth IEEE International and industry in the USA. He was the program co-chair for the 1999 Pacific
Conference on Peer-to-Peer Computing. Washington, DC, USA: IEEE Rim International Symposium on Dependable Computing held in Hong Kong
Computer Society, 2006, pp. 23–30. in December 1999. He also co-founded and organized The Asia-Pacific
[58] J. Pouwelse, P. Garbacki, J. Wang, A. Bakker, J. Yang, A. Iosup, Workshop on Embedded System Education and Research. He has also served
D. Epema, M. Reinders, M. van Steen, and H. Sips, “Tribler: A social- on program committees of many international conferences. He received
based peer-to-peer system,” Concurrency and Computation: Practice & the Excellence in Teaching Innovation Award 2007. He was also awarded
Experience, vol. 20, no. 2, pp. 127–138, Feb 2008. the Teaching Excellence Appreciation Award by the Dean of Engineering,
[59] J. Wang, C. Yeo, V. Prabhakaran, and K. Ramchandran, “On the role HKUST. Dr. Muppala is a Senior Member of IEEE, IEEE Computer Society
of helpers in peer-to-peer file download systems: Design, analysis and and IEEE Communications Society, and a participating representative from
simulation,” in IPTPS’07, 2007. HKUST with EDUCAUSE.

Authorized licensed use limited to: UNIVERSITAT OBERTA DE CATALUNYA. Downloaded on November 21,2023 at 09:02:53 UTC from IEEE Xplore. Restrictions apply.