From pjk25@cornell.edu Mon Apr 3 22:35:39 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-1.7 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from authusersmtp.mail.cornell.edu (granite1.mail.cornell.edu [128.253.83.141]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k342Zdt20702 for ; Mon, 3 Apr 2006 22:35:39 -0400 (EDT) Received: from [10.0.1.3] (cpe-69-207-37-155.twcny.res.rr.com [69.207.37.155]) (authenticated bits=0) by authusersmtp.mail.cornell.edu (8.13.1/8.12.10) with ESMTP id k342ZcmW004307 (version=TLSv1/SSLv3 cipher=RC4-SHA bits=128 verify=NOT) for ; Mon, 3 Apr 2006 22:35:38 -0400 (EDT) Resent-Message-Id: <3D502BD0-7736-4C36-B753-63FDBA7A3ECA@cornell.edu> Mime-Version: 1.0 (Apple Message framework v749.3) Content-Type: text/plain; charset=US-ASCII; delsp=yes; format=flowed Resent-Date: Mon, 3 Apr 2006 22:35:40 -0400 Message-Id: <0184F2D9-1873-4AAC-9E5F-80081230AFA8@cornell.edu> Content-Transfer-Encoding: 7bit Resent-To: egs@cs.cornell.edu From: Philip Kuryloski Subject: PAPER 18 Resent-From: Philip Kuryloski Date: Mon, 3 Apr 2006 16:49:44 -0400 To: egs+summary@cs.cornell.edu X-Mailer: Apple Mail (2.749.3) MODELING AND PERFORMANCE ANALYSIS OF BITTORRENT-LIKE PEER-TO-PEER NETWORKS The paper provides an analysis of the BitTorrent network based on four critical factors: Peer Evolution, Scalability, File Sharing Efficiency, and Incentives to prevent free-riding. Peer Evolution refers to arrival/departure and uploading/downloading bandwidth, etc., of peers over time. File Sharing Efficiency refers to the system's ability to maximize bandwidth utilization across all peers. The authors use a fluid model to analyze the system, and are able to draw several conclusions. The first is that the time required to download a file is independent of the number of requests made for that file, indicating that BitTorrent has excellent scalability. As expected, the greater the file sharing efficiency, the faster the download times, and the faster seeds leave the system, the longer the download times. They find that it if users' download bandwidth is great enough, upload bandwidth becomes the limiting resource slowing down file transfer speeds. However, if seeds leave the system slowly, then downloading bandwidth is the limiting resource of the system. They provide an analysis of file sharing efficiency, and find that because files are typically broken up into several hundred chunks, efficiency is high even when nodes share between very few nodes. Sharing between more nodes increases efficiency, although slowly. Their analysis of incentives leads to several findings. The first is rational users of the system will in fact share files using all available bandwidth, as this will maximize their download speeds. However, due to optimistic unchoking, a free-riding peer will achieve 20% of the maximum downloading rate. The authors obtained experimental results by sharing one of their files. It was downloaded less than 100 times. Traces from this experiment did however match their model to some degree. They also verified other aspects of their model by examining the log file from a BitTorrent tracker. The degree to which the model actually matches the experimental data is somewhat questionable. Visually, one can see that the only plot where model and data match closely is the evolution of the number of seeds. BitTorrent yields a tremendously complex network, due the strong relationship between user behavior and sharing success rates. While the model follows intuition about the network, it does not capture any of the fine grained problems that occur in the network, such as when particular pieces of a file become particularly rare. Also, their data does not come from a popular file, as so may not be representative of a typical scenario. They also seem to assume that users do not change the default settings for their client, without really justifying this. From kelvinso@cs.cornell.edu Tue Apr 4 11:46:34 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-3.2 required=5.0 tests=ALL_TRUSTED,AWL,BAYES_00, HTML_MESSAGE autolearn=ham version=3.1.0 X-Spam-Level: Received: from exchfe1.cs.cornell.edu (exchfe1.cs.cornell.edu [128.84.97.27]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k34FkXt13369 for ; Tue, 4 Apr 2006 11:46:33 -0400 (EDT) Received: from EXCHVS2.cs.cornell.edu ([128.84.97.24]) by exchfe1.cs.cornell.edu with Microsoft SMTPSVC(6.0.3790.1830); Tue, 4 Apr 2006 11:41:41 -0400 x-mimeole: Produced By Microsoft Exchange V6.5 Content-class: urn:content-classes:message MIME-Version: 1.0 X-Security: message sanitized on sundial.cs.cornell.edu See http://www.impsec.org/email-tools/sanitizer-intro.html for details. $Revision: 1.148 $Date: 2004-12-19 11:59:17-08 X-Security: The postmaster has not enabled quarantine of poisoned messages. Content-Type: multipart/alternative; boundary="----_=_NextPart_001_01C657FE.44B45B3B" Subject: Paper 18 Date: Tue, 4 Apr 2006 11:41:41 -0400 Message-ID: <2AA53C9C489B0049B43E56D28088677B2B58EA@EXCHVS2.cs.cornell.edu> X-MS-Has-Attach: X-MS-TNEF-Correlator: Thread-Topic: Paper 18 Thread-Index: AcZX/kSvZzp8l3mNRj6N3jb792SbWA== From: "Kelvin So" To: X-OriginalArrivalTime: 04 Apr 2006 15:41:41.0802 (UTC) FILETIME=[4506C8A0:01C657FE] This is a multi-part message in MIME format. ------_=_NextPart_001_01C657FE.44B45B3B Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable The paper, "Modeling and Performance Anlysis of BitTorrent-Like Peer-to-Peer Networks," presents a model to study the performance of BitTorrent. These paper focuses on several issues in BitTorrent including peer evolution, scalability, file sharing efficiency and incentive to prevent free-riding. Peer evolution is the number of peers in the system over time. Scalability focus on the download rate as the size of the network grows. File sharing efficiency means the efficiency in fully utilizing the bandwidth of all peers. Finally, it looks at how BT uses optimistic unchoking to prevent free-riders in the network. It uses a simple fluid model to capture the performance of = BitTorrent using parameters, such as arrival rate of new requests, uploading and downloading bandwidth of peers, rate of downloader aborts, rate of seeds leave the system and effectiveness of file sharing. It analytical derives the average downloading time, the number of downloaders and seeds in the system over time. Using this equation, it gives some insights to BitTorrent network. One important insight is that the download time is independent on the number of peers. However, in this model it assumes that all peers have uniform upload and download bandwidth. And download bandwidth is always larger than the upload bandwidth. It also shows that rationale peer will share all the available uploading bandwidth when each peer knows the uploading bandwidth of all other peers. When there is limited information about the peers, free-riders will only get 20% of possible maximum downloading rate. Finally, it uses experiments to show the model reflects the behavior of BitTorrent. However, we can see that the results in the experiments are not accurate when the arrival rate is small (when lamda =3D 0.04.) Second, when it compares the result with the real trace, one can tune so many parameters to fit the curve with the real trace to fit the oscillation in the real trace. ------_=_NextPart_001_01C657FE.44B45B3B Content-Type: text/html; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Paper 18

The paper, "Modeling and Performance Anlysis of = BitTorrent-Like
Peer-to-Peer Networks," presents a model to study the performance = of
BitTorrent. These paper focuses on several issues in BitTorrent
including peer evolution, scalability, file sharing efficiency and
incentive to prevent free-riding. Peer evolution is the number of
peers in the system over time. Scalability focus on the download = rate
as the size of the network grows. File sharing efficiency means the
efficiency in fully utilizing the bandwidth of all peers. Finally, = it
looks at how BT uses optimistic unchoking to prevent free-riders in
the network.
       It uses a simple fluid model to = capture the performance of BitTorrent
using parameters, such as arrival rate of new requests, uploading = and
downloading bandwidth of peers, rate of downloader aborts, rate of
seeds leave the system and effectiveness of file sharing. It
analytical derives the average downloading time, the number of
downloaders and seeds in the system over time. Using this equation, = it
gives some insights to BitTorrent network. One important insight is
that the download time is independent on the number of peers. = However,
in this model it assumes that all peers have uniform upload and
download bandwidth. And download bandwidth is always larger than the
upload bandwidth.
       It also shows that rationale peer = will share all the available
uploading bandwidth when each peer knows the uploading bandwidth of
all other peers. When there is limited information about the peers,
free-riders will only get 20% of possible maximum downloading rate.
Finally, it uses experiments to show the model reflects the behavior
of BitTorrent. However, we can see that the results in the = experiments
are not accurate when the arrival rate is small (when lamda =3D = 0.04.)
Second, when it compares the result with the real trace, one can = tune
so many parameters to fit the curve with the real trace to fit the
oscillation in the real trace.

------_=_NextPart_001_01C657FE.44B45B3B-- From asg46@cornell.edu Tue Apr 4 12:12:36 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-2.2 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from postoffice10.mail.cornell.edu (postoffice10.mail.cornell.edu [132.236.56.14]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k34GCZt20718 for ; Tue, 4 Apr 2006 12:12:35 -0400 (EDT) Received: from webmail.cornell.edu (hermes21.mail.cornell.edu [132.236.56.20]) by postoffice10.mail.cornell.edu (8.12.10/8.12.6) with ESMTP id k34GCVFg009332 for ; Tue, 4 Apr 2006 12:12:32 -0400 (EDT) Received: from 128.84.98.90 by webmail.cornell.edu with HTTP; Tue, 4 Apr 2006 12:12:33 -0400 (EDT) Message-ID: <3934.128.84.98.90.1144167153.squirrel@webmail.cornell.edu> Date: Tue, 4 Apr 2006 12:12:33 -0400 (EDT) Subject: paper 18 - unable to send From: "Abhishek Santosh Gupta" To: egs@cs.cornell.edu User-Agent: SquirrelMail/1.4.5 MIME-Version: 1.0 X-Security: message sanitized on sundial.cs.cornell.edu See http://www.impsec.org/email-tools/sanitizer-intro.html for details. $Revision: 1.148 $Date: 2004-12-19 11:59:17-08 X-Security: The postmaster has not enabled quarantine of poisoned messages. Content-Type: multipart/mixed;boundary="----=_20060404121233_81739" X-Priority: 3 (Normal) Importance: Normal ------=_20060404121233_81739 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 8bit ---------------------------- Original Message ---------------------------- Subject: Returned mail: see transcript for details From: "Mail Delivery Subsystem" Date: Tue, April 4, 2006 12:03 am To: asg46@cornell.edu -------------------------------------------------------------------------- The original message was received at Tue, 4 Apr 2006 00:02:55 -0400 (EDT) from hermes21.mail.cornell.edu [132.236.56.20] ----- The following addresses had permanent fatal errors ----- (reason: 550 5.1.1 Bad destination mailbox address (egs+summary@cs.cornell.edu).) ----- Transcript of session follows ----- ... while talking to iago.cs.cornell.edu.: >>> DATA <<< 550 5.1.1 Bad destination mailbox address (egs+summary@cs.cornell.edu). 550 5.1.1 ... User unknown <<< 503 5.5.2 Need Rcpt command. ------=_20060404121233_81739 Content-Type: message/delivery-status; name="untitled-2" Content-Disposition: attachment; filename="untitled-2" Content-Transfer-Encoding: 8bit Reporting-MTA: dns; postoffice10.mail.cornell.edu Received-From-MTA: DNS; hermes21.mail.cornell.edu Arrival-Date: Tue, 4 Apr 2006 00:02:55 -0400 (EDT) Final-Recipient: RFC822; egs+summary@cs.cornell.edu Action: failed Status: 5.1.1 Remote-MTA: DNS; iago.cs.cornell.edu Diagnostic-Code: SMTP; 550 5.1.1 Bad destination mailbox address (egs+summary@cs.cornell.edu). Last-Attempt-Date: Tue, 4 Apr 2006 00:03:57 -0400 (EDT) ------=_20060404121233_81739 Content-Type: message/rfc822; name="paper 18 - MODELING AND PERFORMANCE OF BIT-TORRENT.msg" Content-Disposition: attachment; filename="paper 18 - MODELING AND PERFORMANCE OF BIT-TORRENT.msg" Content-Transfer-Encoding: 8bit Return-Path: Received: from webmail.cornell.edu (hermes21.mail.cornell.edu [132.236.56.20]) by postoffice10.mail.cornell.edu (8.12.10/8.12.6) with ESMTP id k3442tiL001611 for ; Tue, 4 Apr 2006 00:02:55 -0400 (EDT) Received: from 128.84.98.90 by webmail.cornell.edu with HTTP; Tue, 4 Apr 2006 00:02:55 -0400 (EDT) Message-ID: <3254.128.84.98.90.1144123375.squirrel@webmail.cornell.edu> Date: Tue, 4 Apr 2006 00:02:55 -0400 (EDT) Subject: paper 18 - MODELING AND PERFORMANCE OF BIT-TORRENT From: "Abhishek Santosh Gupta" To: egs+summary@cs.cornell.edu User-Agent: SquirrelMail/1.4.5 MIME-Version: 1.0 Content-Type: text/plain;charset=iso-8859-1 Content-Transfer-Encoding: 8bit X-Priority: 3 (Normal) Importance: Normal MODELING AND PERFORMANCE ANALYSIS OF BIT TORRENT.. the paper develops a simple fluid model to study the performance of Bit Torrent. the authors describes Bit Torrent briefly by introducing terminology such as trackers,downloaders,seeders, and optimistic unchoking. FLUID MODEL a number of variables are used to model a number of system parameters. Request arrival is assumed to follow Poisson distribution which is justified by the fact that the number of requests in 2 disjoint intervals is independent. the authors derive mathematical expressions for the following: 1) the total uploading rate taken into account the case when the downloading bandwidth is the constraint 2) rate of departures of downloaders they calculate system parameters in steady state i.e. rate of change of downloaders and seeders each equal to zero. Two cases arise here: 1) downloading bandwidth is the constraint 2) uploading bandwidth is the constraint INSIGHTS: 1) the average downloading time is independent of the arrival rate of requests -- indicating that the system will scale 2) initially when the downloading bandwidth of a peer increases, the average downloading time decreases (obviously). however, after a certain extent, increasing the download bandwidth does not decrease average download time since the bottleneck shifts. the authors also discuss the peer selection algorithm. they conclude that there is not a Nash Equilibrium for a general network setting but one exists when the network consists of groups of peers where members of the group have the same upload and download bandwidths. they point out that optimistic unchoking has a side-effect of encouraging free-riding. ------=_20060404121233_81739-- From sh366@cornell.edu Tue Apr 4 12:34:58 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-2.3 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from postoffice10.mail.cornell.edu (postoffice10.mail.cornell.edu [132.236.56.14]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k34GYwt26150 for ; Tue, 4 Apr 2006 12:34:58 -0400 (EDT) Received: from orpheus3.dataserver.cornell.edu (orpheus3.dataserver.cornell.edu [128.253.161.167]) by postoffice10.mail.cornell.edu (8.12.10/8.12.6) with ESMTP id k34GYuou025069 for ; Tue, 4 Apr 2006 12:34:56 -0400 (EDT) Message-ID: <1707721758.1144168496139.JavaMail.webber@orpheus3.dataserver.cornell.edu> Date: Tue, 4 Apr 2006 12:34:56 -0400 (EDT) From: Huang Shiang-Jia To: egs+summary@cs.cornell.edu Subject: PAPER 18 Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit X-Mailer: uPortal WEB email client 3.0 This paper presents a fluid model to study the scalability, performance and efficiency of BitTorrent-like systems and discusses the incentive mechanism in BitTorrent to discourage free-riders. * The fluid model in this paper expresses the average number of seeds (nodes that have all pieces of files for others' downloads), average number of downloaders, and average download time as functions of downloader arrival/leaving rates, upload/download bandwidths and seed leaving rates. * It shows that: (a) the download time is irrelevant to peer arrival rates, thus the system is scalable; (b) the download time increases if many seeds leave the system; (c) the download bandwidth is no longer the bottleneck of the download time if it's large enough. * It then shows that the effectiveness (uploading files to others) of downloaders play an important role to keep the system alive. As long as downloaders upload/share files (even though only a few) with others, the system will eventually reach a steady state. * It also shows that the effectiveness g (which takes values in [0, 1]) is very close to 1 in BitTorrent. Because the value of g is relevant to k, the number of peers a downloader connects to, which in turn is related to the peer arrival rates, this paper concludes that the performance of BitTorrent improves as the size of the system increases. * It finally shows that, affected by the peer selection rules, each peer chooses its uploading bandwidth equal to the actual uploading bandwidth; that is, Nash equilibrium exists. This paper shows the effect of optimistic choking on free-riders: free-riders gets an average download rate in inverse proportion to n+1, where n is the number of a peer's uploads (4 in current BitTorrent; in this case, download rate of a free-rider is about 20%). To decide an optimal n to balance system performance with deterrence of free-riders is an open issue. From victoria@cs.hmc.edu Tue Apr 4 12:44:35 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-2.2 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from iago.cs.cornell.edu (iago.cs.cornell.edu [128.84.96.10]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k34GiYt28231 for ; Tue, 4 Apr 2006 12:44:34 -0400 (EDT) Received: from turing.cs.hmc.edu ([134.173.42.99]) by iago.cs.cornell.edu with Microsoft SMTPSVC(6.0.3790.1830); Tue, 4 Apr 2006 12:41:42 -0400 Received: by turing.cs.hmc.edu (Postfix, from userid 34382) id 6E28353245; Tue, 4 Apr 2006 09:23:41 -0700 (PDT) Date: Tue, 4 Apr 2006 09:23:41 -0700 From: Victoria Krafft To: egs+summary@cs.cornell.edu Subject: PAPER 18 Message-ID: <20060404162341.GA14758@cs.hmc.edu> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline User-Agent: Mutt/1.4.2.1i X-OriginalArrivalTime: 04 Apr 2006 16:41:42.0729 (UTC) FILETIME=[A758A390:01C65806] This paper presents a mathematical model for the behavior of a BitTorrent network. Their model explains several features of the BitTorrent network. In their model, the average downloading time is independent of the request arrival rate, which explains why BitTorrent can handle flash crowds. The average downloading time is related to both the effectiveness of the file sharing, and then rate at which seeds leave the system. They go on to present a model for the effectiveness of filesharing in BitTorrent, and argue that it will always be fairly close to 1. Their analysis of the incentive mechanism in BitTorrent suggests that in a reasonably large network, a freeloading peer will get about 20% of the download bandwidth a well-behaved peer receives. The comparison of the predicted behavior of a BitTorrent network under this model, and an actual BitTorrent network, is extremely limited. Most of the comparison is done between their model and a simulated BitTorrent network. While the general trends in these simulations match up with the model, the simulations have some large oscillations which the model does not predict. The lack of error bars for the simulation data makes it difficult to determine if this is caused by random variation within the simulations, or if the model is failing to predict some behavior. The larger the arrival rate of new requests is, the more accurate the model appears to be. At the end, there is a comparison to the behavior of an actual BitTorrent network. However, the limited popularity of this file makes it difficult to draw any conclusions about the accuracy of the model. While the evolution of the number of seeds and number of downloaders is within the 95% confidence interval, that interval is rather large, especially for the number of downloaders. While having a model of the behavior of BitTorrent would be useful, I'm not convinced that the model presented here will accurately predict the behavior of real BitTorrent networks. -- Victoria Krafft From gp72@cornell.edu Tue Apr 4 12:46:50 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-2.0 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from postoffice10.mail.cornell.edu (postoffice10.mail.cornell.edu [132.236.56.14]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k34Gknt28585 for ; Tue, 4 Apr 2006 12:46:49 -0400 (EDT) Received: from orpheus3.dataserver.cornell.edu (orpheus3.dataserver.cornell.edu [128.253.161.167]) by postoffice10.mail.cornell.edu (8.12.10/8.12.6) with ESMTP id k34Gkm6B003425 for ; Tue, 4 Apr 2006 12:46:48 -0400 (EDT) Message-ID: <1884611382.1144169207365.JavaMail.webber@orpheus3.dataserver.cornell.edu> Date: Tue, 4 Apr 2006 12:46:47 -0400 (EDT) From: Gopal Parameswaran To: egs+summary@cs.cornell.edu Subject: PAPER 18 Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii X-Mailer: uPortal WEB email client 3.0 Content-Transfer-Encoding: 8bit X-MIME-Autoconverted: from quoted-printable to 8bit by sundial.cs.cornell.edu id k34Gknt28585 -----Original Message----- > Date: Tue Apr 04 11:48:07 EDT 2006 > From: "Gopal Parameswaran" > Subject: PAPER 18 > To: egs+summary@cs.cornell.edu > > Modeling and Performance Analysis of BitTorrent-Like Peer-to-Peer Networks > This paper develop a simple fluid model to study the performance, scalability, performance and efficiency of Bit Torrent, a file-sharing mechanism of a second generation peer-to-peer (P2P) application. A numerical analysis is performed and numerical results based on both simulations and real traces obtained from the Internet are provided. In more specificity the number of peers in the system and its effect in bit torrent along with the scalability, file sharing efficiency when upload and download speeds are different and the incentives for free riding. The authors develop a develop a simple deterministic model to obtain simple expressions for the average file-transfer time, thus providing insight into the performance of the P2P network with realistic scenarios in the fluid model such as the abandonment of file transfers by peers and download bandwidth constraints and a simple stochastic fluid model which characterizes the variability of the number of peer around the equilibrium values predicted by the deterministic fluid model. This paper obtains expressions for the average number of seeds, the average number of downloader, and the average downloading time as functions of the peer arrival rate, downloader leaving rate, seed leaving rate, uploading bandwidth, etc, which explicitly give us insight on how the network performance is affected by different parameters. This paper also characterized the variability of the system by applying limit theorems to the stochastic model when the arrival rate is large. This paper also discusses the effect of optimistic unchoking on free-riding and how the simple fluid model can capture the behavior of the system even when the arrival rate is small. From asr32@cornell.edu Wed Apr 5 00:30:09 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-2.0 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from authusersmtp.mail.cornell.edu (granite1.mail.cornell.edu [128.253.83.141]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k354U8t29277 for ; Wed, 5 Apr 2006 00:30:08 -0400 (EDT) Received: from dreadnought.cornell.edu (r253240123.resnet.cornell.edu [128.253.240.123]) (authenticated bits=0) by authusersmtp.mail.cornell.edu (8.13.1/8.12.10) with ESMTP id k354U7Ms002341 (version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NOT) for ; Wed, 5 Apr 2006 00:30:08 -0400 (EDT) Message-Id: <6.2.1.2.2.20060405003021.03413468@postoffice8.mail.cornell.edu> X-Mailer: QUALCOMM Windows Eudora Version 6.2.1.2 Date: Wed, 05 Apr 2006 00:30:31 -0400 To: egs+summary@cs.cornell.edu From: Ari Rabkin Subject: PAPER 18 Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii"; format=flowed [Sending late due to submission troubles] Analysis of BitTorrent: The design of BitTorrent was very much ad-hoc. Here, the authors seek to build a rigorous model of the network, using markov processes. They are able to derive plausible estimates for probability of downloads succeeding, and for the population sizes. They demonstrate that under reasonable assumptions, the system converges to a Nash equilibrium where the optimum strategy for every node is to upload. That is, the BitTorrent incentive system seems to work. Actual BitTorrent networks have a lot of short-term variation; the graphs of real networks are jagged, and the authors' predictions are smooth. The model does not take short-term variation into account, and as a result, real BitTorrent networks can significantly deviate from optimum. The proof that the system will converge to a Nash equilibrium seems predicated on global knowledge; there is no reason why a real BitTorrent like system will so converge. Ari Rabkin asr32@cornell.edu Risley Hall 454 3-2842 The resources of civilization are not yet exhausted. --William Gladstone From nsg7@cornell.edu Wed Apr 5 09:21:51 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-2.1 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from postoffice10.mail.cornell.edu (postoffice10.mail.cornell.edu [132.236.56.14]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k35DLpt23102 for ; Wed, 5 Apr 2006 09:21:51 -0400 (EDT) Received: from webmail.cornell.edu (hermes21.mail.cornell.edu [132.236.56.20]) by postoffice10.mail.cornell.edu (8.12.10/8.12.6) with ESMTP id k35DLmw7003620 for ; Wed, 5 Apr 2006 09:21:49 -0400 (EDT) Received: from 132.236.227.192 by webmail.cornell.edu with HTTP; Wed, 5 Apr 2006 09:21:49 -0400 (EDT) Message-ID: <1431.132.236.227.192.1144243309.squirrel@webmail.cornell.edu> Date: Wed, 5 Apr 2006 09:21:49 -0400 (EDT) Subject: PAPER 18 From: "Nicholas S Gerner" To: egs+summary@cs.cornell.edu User-Agent: SquirrelMail/1.4.5 MIME-Version: 1.0 Content-Type: text/plain;charset=iso-8859-1 Content-Transfer-Encoding: 8bit X-Priority: 3 (Normal) Importance: Normal Qiu and Srikant present an analysis of BitTorrent like networks in "Modeling and Performance Analysis of BitTorrent-Like Peer-to-Peer Networks". "BitTorrent like networks" means that files are divided into chunks which are downloaded from seeds. Additionally downloaders can download chunks from each other even if they don't have the entire file. Qiu and Srikant develop a model to explore peer evolution, scalability, file sharing efficiency and incentives to participate. This is a fluid model which captures the number of seeds, downloaders, arrival rate of downloaders, uploading and downloading bandwidth, rate at which downloaders and seeds leave the system. An important aspect of the system is the efficiency which is characterized by eta ranging from 0 to 1. At eta=0 downloaders do not upload. At eta=1 all downloaders are uploading to each other. This is an important aspect because this is an important contribution of BitTorrent: that downloaders are able to improve system performance because they can be uploaders as they download the files. Qiu and Srikant show that in BitTorrent, if all peers follow the protocol, eta approaches 1 with N and the number of uploading connections. The model is also used to show that the uploading bandwidth of the system grows with the number of downloaders and in fact the download bandwidth of each peer is independent of the number of downloaders and so the system is very scalable. Another important contribution of the paper is an analysis of the Nash equilibrium of agents in such a system. In fact, if optimistic unchoking is ignored, each peer would upload as much as possible in order to maximize downloading (due to incentives in the BitTorrent protocol). However, optimistic unchoking could lead to free-riders (since it allows peers to download without uploading), but further analysis shows that this will only utilize a fraction (small given BitTorrent parameters) of the bandwidth possible if peers do upload. This analysis suggests that BitTorrent is efficient, scalable and provides good incentives for participation (if maximizing download bandwidth is more important than minimizing upload bandwidth). Simulation shows that in many cases the model accurately characterizes the number of seeds and downloaders (for large downloader arrival rates). An analysis of a real trace shows that some significant deviation from the model exists. Qiu and Srikant argue that this deviation is within expected 95% confidence intervals from the predicted values of number of seeds and downloaders. Further, they argue that the file they considered was unpopular and therefore did not exhibit high enough downloder arrival rates to fit the model well. Qiu and Srikant also note that while they characterized downloader arrival and seed departure by poisson processes, a constant parameter for these distributions was not observed in a real trace. From pjk25@cornell.edu Thu Apr 6 01:00:53 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-1.9 required=5.0 tests=AWL,BAYES_00 autolearn=unavailable version=3.1.0 X-Spam-Level: Received: from authusersmtp.mail.cornell.edu (granite1.mail.cornell.edu [128.253.83.141]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k3650qt19511 for ; Thu, 6 Apr 2006 01:00:52 -0400 (EDT) Received: from [10.0.1.3] (cpe-69-207-37-155.twcny.res.rr.com [69.207.37.155]) (authenticated bits=0) by authusersmtp.mail.cornell.edu (8.13.1/8.12.10) with ESMTP id k3650pfR007658 (version=TLSv1/SSLv3 cipher=RC4-SHA bits=128 verify=NOT) for ; Thu, 6 Apr 2006 01:00:52 -0400 (EDT) Resent-Message-Id: <0184F2D9-1873-4AAC-9E5F-80081230AFA8@cornell.edu> Mime-Version: 1.0 (Apple Message framework v749.3) Content-Type: text/plain; charset=US-ASCII; delsp=yes; format=flowed Resent-Date: Thu, 6 Apr 2006 01:00:54 -0400 Message-Id: <186451A7-5317-4271-A6D3-B942FB4CCEAC@cornell.edu> Content-Transfer-Encoding: 7bit Resent-To: egs+summary@cs.cornell.edu From: Philip Kuryloski Subject: PAPER 18 Resent-From: Philip Kuryloski Date: Mon, 3 Apr 2006 16:49:44 -0400 To: egs+summary@cs.cornell.edu X-Mailer: Apple Mail (2.749.3) MODELING AND PERFORMANCE ANALYSIS OF BITTORRENT-LIKE PEER-TO-PEER NETWORKS The paper provides an analysis of the BitTorrent network based on four critical factors: Peer Evolution, Scalability, File Sharing Efficiency, and Incentives to prevent free-riding. Peer Evolution refers to arrival/departure and uploading/downloading bandwidth, etc., of peers over time. File Sharing Efficiency refers to the system's ability to maximize bandwidth utilization across all peers. The authors use a fluid model to analyze the system, and are able to draw several conclusions. The first is that the time required to download a file is independent of the number of requests made for that file, indicating that BitTorrent has excellent scalability. As expected, the greater the file sharing efficiency, the faster the download times, and the faster seeds leave the system, the longer the download times. They find that it if users' download bandwidth is great enough, upload bandwidth becomes the limiting resource slowing down file transfer speeds. However, if seeds leave the system slowly, then downloading bandwidth is the limiting resource of the system. They provide an analysis of file sharing efficiency, and find that because files are typically broken up into several hundred chunks, efficiency is high even when nodes share between very few nodes. Sharing between more nodes increases efficiency, although slowly. Their analysis of incentives leads to several findings. The first is rational users of the system will in fact share files using all available bandwidth, as this will maximize their download speeds. However, due to optimistic unchoking, a free-riding peer will achieve 20% of the maximum downloading rate. The authors obtained experimental results by sharing one of their files. It was downloaded less than 100 times. Traces from this experiment did however match their model to some degree. They also verified other aspects of their model by examining the log file from a BitTorrent tracker. The degree to which the model actually matches the experimental data is somewhat questionable. Visually, one can see that the only plot where model and data match closely is the evolution of the number of seeds. BitTorrent yields a tremendously complex network, due the strong relationship between user behavior and sharing success rates. While the model follows intuition about the network, it does not capture any of the fine grained problems that occur in the network, such as when particular pieces of a file become particularly rare. Also, their data does not come from a popular file, as so may not be representative of a typical scenario. They also seem to assume that users do not change the default settings for their client, without really justifying this. From niranjan.sivakumar@gmail.com Thu Apr 6 02:32:05 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-1.7 required=5.0 tests=AWL,BAYES_00,HTML_00_10, HTML_MESSAGE autolearn=no version=3.1.0 X-Spam-Level: Received: from penguin.cs.cornell.edu (penguin.cs.cornell.edu [128.84.96.11]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.22) with ESMTP id k366W5t07946 for ; Thu, 6 Apr 2006 02:32:05 -0400 (EDT) Received: from xproxy.gmail.com ([66.249.82.194]) by penguin.cs.cornell.edu with Microsoft SMTPSVC(6.0.3790.1830); Thu, 6 Apr 2006 02:32:05 -0400 Received: by xproxy.gmail.com with SMTP id t11so51177wxc for ; Wed, 05 Apr 2006 23:32:05 -0700 (PDT) DomainKey-Signature: a=rsa-sha1; q=dns; c=nofws; s=beta; d=gmail.com; h=received:message-id:date:from:to:subject:in-reply-to:mime-version:content-type:references; b=uO/FXgdr3OJB0Ir2SH/UJRsaC2itEvBONTsaYSjAUvidMLg9vrS5LGaANL2efNNfFuOTgCS0qCeMGulzuJTs/xB/9a8xFziT7YqpscMtwnrBpD8v1x8cOEKerLHOyhJ7RGzH+x8BQodjJa385dkHKLXDe7YMH3miqmMZXQJ/3Is= Received: by 10.70.97.14 with SMTP id u14mr655294wxb; Wed, 05 Apr 2006 22:38:01 -0700 (PDT) Received: by 10.70.128.19 with HTTP; Wed, 5 Apr 2006 22:38:01 -0700 (PDT) Message-ID: Date: Thu, 6 Apr 2006 01:38:01 -0400 From: "Niranjan Sivakumar" To: egs+summary@cs.cornell.edu Subject: PAPER 18 In-Reply-To: MIME-Version: 1.0 X-Security: message sanitized on sundial.cs.cornell.edu See http://www.impsec.org/email-tools/sanitizer-intro.html for details. $Revision: 1.148 $Date: 2004-12-19 11:59:17-08 X-Security: The postmaster has not enabled quarantine of poisoned messages. Content-Type: multipart/alternative; boundary="----=_Part_21858_32378637.1144301881565" References: X-OriginalArrivalTime: 06 Apr 2006 06:32:05.0557 (UTC) FILETIME=[D27A9E50:01C65943] ------=_Part_21858_32378637.1144301881565 Content-Type: text/plain; charset=ISO-8859-1 Content-Disposition: inline Content-Transfer-Encoding: quoted-printable Niranjan Sivakumar Modeling and Performance Analysis of BitTorrent-Like Peer-to-Peer Networks This paper deals with the development of a model to analyze a variety of performance related characteristics of a BitTorrent. The main issues that were considered are peer evolution, scalability, effectiveness of file sharing, and incentives. The authors based their approach on a fluid model= . Peers in the model are considered to not always "play by the rules" and there are provisions for simulating peers that may leave the system without completing a download and seeding. The authors claim that scalability and file sharing effectiveness are generally good in BitTorrent. The download rate did not seem to be adversely affected by the rate of incoming queries, meaning that the system did not generally get bogged down as more peers joined. As the arrival rate of peers increases, performance should slowly improve. An interesting observation that was made is that although upload speeds may often be much slower than download speeds for Internet connections, the upload rate was not always the limiting factor in BitTorrent swarms. In cases where the seed departure rate was less than th= e joining rate, the downloding bandwidth ultimately determined the performance. The incentive mechanism was shown generally to incentivize sharing files at a high rate. However, it was also noted that BitTorrent itself does not provide any incentives for nodes to seed files. Free riders were shown to be able to get 20% of the possible maximum downloading rate because of optimistic unchoking. This was considered to be a problem, and it is proposed that the system may be tuned to decrease it. One of the biggest issues seen with this paper is that a "popular" file was not inserted into the network. The claim that introducing a popular file into a BitTorrent network would necessarily infringe copyrights seems to be a dubious claim that provided for a less than optimal simulation. While there is probably a legitimate debate as to the quantity of files transferred with BitTorrent that may be infringing, there are certainly ver= y visible examples of legitimate, non infringing files and very popular content that has been distributed via the network. ------=_Part_21858_32378637.1144301881565 Content-Type: text/html; charset=ISO-8859-1 Content-Disposition: inline Content-Transfer-Encoding: quoted-printable Niranjan Sivakumar

Modeling and Perfo= rmance Analysis of BitTorrent-Like Peer-to-Peer Networks

This paper = deals with the development of a model to analyze a variety of performance r= elated characteristics of a BitTorrent.  The main issues that were con= sidered are peer evolution, scalability, effectiveness of file sharing, and= incentives.  The authors based their approach on a fluid model.

Peers in the model are considered to not always "play by the r= ules" and there are provisions for simulating peers that may leave the= system without completing a download and seeding.  The authors claim = that scalability and file sharing effectiveness are generally good in BitTo= rrent.  The download rate did not seem to be adversely affected by the= rate of incoming queries, meaning that the system did not generally get bo= gged down as more peers joined.  As the arrival rate of peers increase= s, performance should slowly improve.  An interesting observation that= was made is that although upload speeds may often be much slower than down= load speeds for Internet connections, the upload rate was not always the li= miting factor in BitTorrent swarms.  In cases where the seed departure= rate was less than the joining rate, the downloding bandwidth ultimately d= etermined the performance.  The incentive mechanism was shown generall= y to incentivize sharing files at a high rate.  However, it was also n= oted that BitTorrent itself does not provide any incentives for nodes to se= ed files. Free riders were shown to be able to get 20% of the possible maxi= mum downloading rate because of optimistic unchoking.  This was consid= ered to be a problem, and it is proposed that the system may be tuned to de= crease it.

One of the biggest issues seen with this paper is that a "popu= lar" file was not inserted into the network.  The claim that intr= oducing a popular file into a BitTorrent network would necessarily infringe= copyrights seems to be a dubious claim that provided for a less than optim= al simulation.  While there is probably a legitimate debate as to the = quantity of files transferred with BitTorrent that may be infringing, there= are certainly very visible examples of legitimate, non infringing files an= d very popular content that has been distributed via the network.
------=_Part_21858_32378637.1144301881565-- From asg46@cornell.edu Tue Apr 11 12:07:22 2006 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.1.0 (2005-09-13) on sundial.cs.cornell.edu X-Spam-Status: No, score=-2.2 required=5.0 tests=AWL,BAYES_00 autolearn=ham version=3.1.0 X-Spam-Level: Received: from postoffice10.mail.cornell.edu (postoffice10.mail.cornell.edu [132.236.56.14]) by sundial.cs.cornell.edu (8.11.7-20031020/8.11.7/M-3.25) with ESMTP id k3BG7M212965 for ; Tue, 11 Apr 2006 12:07:22 -0400 (EDT) Received: from webmail.cornell.edu (hermes21.mail.cornell.edu [132.236.56.20]) by postoffice10.mail.cornell.edu (8.12.10/8.12.6) with ESMTP id k3BG7Kvx023751 for ; Tue, 11 Apr 2006 12:07:20 -0400 (EDT) Received: from 128.84.97.226 by webmail.cornell.edu with HTTP; Tue, 11 Apr 2006 12:07:21 -0400 (EDT) Message-ID: <4088.128.84.97.226.1144771641.squirrel@webmail.cornell.edu> Date: Tue, 11 Apr 2006 12:07:21 -0400 (EDT) Subject: paper 18 From: "Abhishek Santosh Gupta" To: egs+summary@cs.cornell.edu User-Agent: SquirrelMail/1.4.5 MIME-Version: 1.0 Content-Type: text/plain;charset=iso-8859-1 Content-Transfer-Encoding: 8bit X-Priority: 3 (Normal) Importance: Normal SCRIBE a scalable application-level multicast infrastructure supporting a large number of groups with any number of members in a group. Scribe is built on the top of Pastry. Pastry is used to create and manage groups and build trees for efficient multicast distribution. each group has a unique group-id which is computed using the hash of the textual name concatenated with the creator's name. the scribe node with the nodeid closest to the group-id acts as the root of the multi-cast tree for the group. the authors also propose making the creator the root of the multicast tree which seems like a better idea - this prevents malicious nodes from overburdening other nodes by creating groups with group-id that are close to the nodeid of the correct node. when a scribe node joins, it uses the Pastry's join message along with the group-id that it wishes to join. this message is routed using Pastry. a node is termed as a forwarder for a group if its routing table contains entries for that group. A node becomes a forwarder for a group when a node sends it a JOIN message for a particular group (which is not in its routing table and provided its routing table is free)\ the root of a group is located using Pastry and also cached for subsequent use. multicast trees repair: parents send heartbeats to their children periodically. when a child detects failure of its parent it sends a JOIN message. root failures can be tolerated using replication and agreement protocols. DESIGN AND EVALUATION ...:SIENA the authors discuss the tradeoff between the maximizing the expressiveness in the selection mechanism and maintain scalability at the same time. the selection process is used to select clients while routing and delivering notifications. for a distributed system implementing a notification system, three critical design aspects are mentioned: 1) interconnection topology between servers 2) routing algorithm 3) processing strategy -- so as to optimize message traffic the authors adopt a general graph for their interconnection topology based on performance issues with a hierarchal topology. besides the publish/subscribe interface, the authors also use an additional interface termed advertise which an object of interest uses to advertise notifications it publishes ( unsuscribe and unadvertise also exist) SIENA semantics are that of a best-effort service - it must not introduce unnecessary delays in processing but it neither prevents race conditions. a timestamp with each notification is used to deal with latency effects. the authors propose a hybrid architecture so as to increase efficiency by using different architectures at different points in the network. Patterns are matched using "Pattern Factoring" and "Pattern Delegation" mechanisms which break a compound subscription into components and forward the smaller elementary components.