From eyh5@ee.cornell.edu Wed Dec 5 18:03:07 2001 Received: from memphis.ece.cornell.edu (memphis.ece.cornell.edu [128.84.81.8]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB5N37608365 for ; Wed, 5 Dec 2001 18:03:07 -0500 (EST) Received: from photon.ece.cornell.edu (photon.ece.cornell.edu [128.84.81.138]) by memphis.ece.cornell.edu (8.11.6/8.11.2) with ESMTP id fB5MxMM26504 for ; Wed, 5 Dec 2001 17:59:22 -0500 Date: Wed, 5 Dec 2001 17:59:23 -0500 (EST) From: Edward Hua X-X-Sender: To: Subject: 615 Paper # 76 Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII A Measurement Study of Peer-to-Peer File Sharing Systems Stefan Saroiu, P. Krishna Gummadi, Steven D. Gribble This paper offers a qualitative evaluation of two of the more popularly known peer-to-peer file sharing systems, Napster and Gnutella. The need for such an evaluation arises from the perceived fact that, while a number of proposed peer-to-peer systems have been touted in recent years, few are accompanied by qualified evaluations that examine the implications of suitably choosing peer node to serve the functionality required by such a peer-to-peer network. The paper seeks to characterize the population of end-user hosts in these two systems from a number of aspects, including the bootleneck bandwidths between hosts and Internet, IP-level latencies, duration of host remaining connected to the network, amount of files processed by the host, and correlations of the above. The technique used to evaluate Napster and Gnutella in this paper is composed of two steps. First, a network crawler is developed for each of the two systems to gather snapshot information about the network.In Napster, the network crawler sends queries to a cluster of 160 Napster servers with names of popular song artists, and save the peers that these queries return; then the snapshot information is directly gathered about these peers by querying the servers again. In Gnutella, the crawler uses the ping/pong message to discover the peer hosts and learns of their status information such as bandwidth and the number of files shared. Then, real-time probing is applied to the users captured in the snapshot information to measure various properties and behavioral partterns over a period of time. The probing aims to obtain information such as latency measurements, lifetime measurements, and bottleneck bandwidth measurements. With the data gathered at hand, the technique yields some interesting results about the behavioral patterns in Napster and Gnutella. Two lessons are learned from the evaluation presented in this paper. First any similar peer-to-peer system must be very careful about delegating responsibilities across peers; and second, future systems must have built-in incentivies for peers to behave honestly, or systems must be able directly measure or verify reported information. A few comments, however, seem to be in order here. First, the acquisition and the interpretation of the results seem to be very subjective. Second, the results uncovered in this paper only show how many peer hosts are capable of serving as servers, but not how many are willing to serve as such. Finally, this technique is rather intrusive and may not be so easily applied to evaluate other peer-to-peer networks, as they probably will have more stringent security measures to protect the identities of both the servers and client peers, thus making the acquisition of information such as number of files in a peer host hard to come by. From wbell@CS.Cornell.EDU Wed Dec 5 21:06:55 2001 Received: from postoffice.mail.cornell.edu (postoffice.mail.cornell.edu [132.236.56.7]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB626t606429 for ; Wed, 5 Dec 2001 21:06:55 -0500 (EST) Received: from [192.168.1.100] (syr-66-24-16-64.twcny.rr.com [66.24.16.64]) by postoffice.mail.cornell.edu (8.9.3/8.9.3) with ESMTP id VAA00740 for ; Wed, 5 Dec 2001 21:06:53 -0500 (EST) Subject: 615 PAPER #76 From: Walter Bell To: egs@CS.Cornell.EDU Content-Type: text/plain Content-Transfer-Encoding: 7bit X-Mailer: Evolution/0.99.1+cvs.2001.11.07.16.47 (Preview Release) Date: 05 Dec 2001 21:06:32 -0500 Message-Id: <1007604415.905.0.camel@brute> Mime-Version: 1.0 76) A Measurement Study of Peer-to-Peer File Sharing Systems This paper presents a detailed study of the behaviour and properties of users of the Napster and Gnutella file sharing services, examining a large portion of the overall network over a period of a few days. They examined the number of files shared by users, as well as download and upload statistics for users as well as bandwidth statistics. Their results show what our intuition tells us; users won't share unless forced to, and there is a large heterogeneity in the properties of the hosts in both systems. Users were found to misreport statistics and not share files in order to save network resources which forced a large amount of the network load to be put on a small number of hosts. The heterogeneity in hosts only leads us to believe that the current symmetric views of hosts in peer to peer networks is a bad idea; with nodes that are diversely heterogeneous, we should allocate resources to nodes that are the most suitable for the job, instead of equally across all nodes. I think this report should force people to rethink the views we have on current large scale system design-- as I said, it only presented us with things that our intuition already led us to, but with them presented in a well done study, this work should be reflected in future designs for wide area peer to peer networks. This motivates further the need for configuration less self-organizing systems, because configuration allows users to manipulate their participation in the network to their advantage. From daehyun@csl.cornell.edu Thu Dec 6 11:09:14 2001 Received: from wilkes.csl.cornell.edu (wilkes.csl.cornell.edu [132.236.71.69]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6G9D621118 for ; Thu, 6 Dec 2001 11:09:13 -0500 (EST) Received: (from daehyun@localhost) by wilkes.csl.cornell.edu (8.9.3/8.9.2) id LAA78604 for egs@cs.cornell.edu; Thu, 6 Dec 2001 11:09:08 -0500 (EST) (envelope-from daehyun) From: Daehyun Kim Message-Id: <200112061609.LAA78604@wilkes.csl.cornell.edu> Subject: 615 PAPER 76 To: egs@CS.Cornell.EDU Date: Thu, 6 Dec 2001 11:09:08 -0500 (EST) X-Mailer: ELM [version 2.4ME+ PL54 (25)] MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit This paper presents a detailed measurement study of the two most popular peer-to-peer file sharing systems, Napster and Gnutella. This paper shows precise characterization such as the bottleneck bandwidth between the peers and the Internet at large, IP-level latencies to send packets to the peers, how often the peers connect and disconnect from the system, how many files the peers share and download, the degree of cooperation between the peers, and several correlations between theses characteristics. The methodology used in this paper is two steps. First, they periodically crawled each system to gather instantaneous snapshots of large subsets of the systems' user population. Second, they actively probed the users in the snapshot over a period of several days to directly measure various properties such as their bottleneck bandwidth. In conclusion, this paper says that there is significant heterogeneity and lack of cooperation across the peers. First, Both Napster and Gnutella show heterogeneity in bandwidth, latency, availability, and the degree of sharing between three and five orders of magnitude. So, future systems must be careful about delegating responsibilities across the peers. Second, Peers tend to misreport information if there is an incentive to do so. So future systems must have built-in incentives for the peers to tell the truth, or the systems must be able to directly measure or verify reported information. In my opinion, the measurement study in this paper is very nice attempt. We have read many papers about ad-hoc network and peer-to-peer systems. They say how to build scalable and efficient systems and evaluates their systems by simulation. Though the input parameters such as traffic patterns are absolute factors that decide the results, they usually chose arbitrary parameters, which might results in wrong conclusions. Measurement studies such as this paper will give more practical inputs to simulations, then more precise system evaluations will be possible. From avneesh@csl.cornell.edu Thu Dec 6 11:45:38 2001 Received: from capricorn.ds.csl.cornell.edu (capricorn.csl.cornell.edu [132.236.71.92]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6Gjb627325 for ; Thu, 6 Dec 2001 11:45:37 -0500 (EST) content-class: urn:content-classes:message Subject: 615 Paper 76 MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Date: Thu, 6 Dec 2001 11:49:02 -0500 X-MimeOLE: Produced By Microsoft Exchange V6.0.4712.0 Message-ID: <97C142C1212ED545B0023A177F5349C40A0A33@capricorn.ds.csl.cornell.edu> X-MS-Has-Attach: X-MS-TNEF-Correlator: Thread-Topic: 615 Paper 76 Thread-Index: AcF+deifTT4n9tFCReC8KM05IHPMpw== From: "Avneesh Bhatnagar" To: Content-Transfer-Encoding: 8bit X-MIME-Autoconverted: from quoted-printable to 8bit by sundial.cs.cornell.edu id fB6Gjb627325 A measurement study of peer to peer file systems This paper presents a study of Napster and Gnutella peer to peer share systems. The main factors that were evaluated were: a. Latency/availability measurements. These are quite closely related to the bottleneck bandwidth measurements, which in other words implies the rate at which content can be uploaded/downloaded from a peer. The measurements of this was done using a variation of the SProbe tool. b. Lifetime measurements: These indicate the average uptimes of peers and also correlates to how the network would look like if some of the popular peers went down. In the case of Napster, central server failure is an issue, while in Gnutella disjoint Gnutella groups would be formed. The measurements point out some interesting characteristics and help answer some questions: a.Even though systems might be designed with equal responsiblities in mind, these systems might diverge towards client and server groups, determined to quite a large extent by the quality of the last hop/link. Furthermore tendency of peers to report inaccurate capabilities is also an important issue in the assymetry. b. In such a system, if peer's connections are forged in an adhoc way, then a substantial fraction of connections face a high latency. Also latencies are impacted if the comunication e.g takes place between trans-oceanic peers rather than peers on the same part/opposite part of the continent. c.Napster fits comes out to be a higher quality and more useful servicewith regards to the percentage of peers offering a higher availablity profile. Furthermore Napster peers are slightly more consistent and offer less varition in the number of shared files. These factors could be attributed to the fact that Gnutella has not yet taken off as well as Napster did, since the inherent protocol involves a lot of flooding/neighbor discovery traffic and thus yield overall higher latencies. With the exit of Napster, we might see a change in the files shared trend. d. Gnutella shows high resilience in terms of attacks, but might suffer badly if the attacks are well planned and effect high responsibility peers. I think that this study was very interesting and the authors have pointed one of the concerns I felt would have been common with a system like Tapestry and that is the fact that we might assume very fast and high end peers, but that does not remove smaller capacity peers from the system. This might create vast imbalances in the donwload speeds that are offered as well as the number of files etc. From this study it can be seen that the type of communication mode (cabel/DSL/modem) being used greatly affects the upload/download characteristics. Finally all peers might not cooperate equally in file sharing or may misreport their capabilties. This leads to some peers getting more traffic than others. Finally, the only issue that might impair a study like this, is that it might not be able to design 'crawlers' for all peer to peer systems. Having said this, this study adequately points out a set of attributes that can be used to evaluate the efficacy of any peer to peer system. From viran@csl.cornell.edu Thu Dec 6 12:03:10 2001 Received: from moore.csl.cornell.edu (moore.csl.cornell.edu [132.236.71.83]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6H3A600465 for ; Thu, 6 Dec 2001 12:03:10 -0500 (EST) Received: from localhost (viran@localhost) by moore.csl.cornell.edu (8.11.3/8.9.2) with ESMTP id fB6H34N35025 for ; Thu, 6 Dec 2001 12:03:04 -0500 (EST) (envelope-from viran@moore.csl.cornell.edu) X-Authentication-Warning: moore.csl.cornell.edu: viran owned process doing -bs Date: Thu, 6 Dec 2001 12:03:04 -0500 (EST) From: "Virantha N. Ekanayake" To: Subject: 615 Paper 76 Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII This paper presents a quantitative performance, usage, and traffic comparison of Gnutella and Napster, and the usage patterns. They did this by getting snapshots of that status of the network by querying the central server in napster, and by using the ping-pong protocol in gnutella. They report bandwidth, latency, download numbers, and file-sharing habits of the users on either network. They also measured how many peers are like servers (the degree of asymmetry), which leads to a measure of the resiliency of the network in question. It's results are presented in a very clear and straightforward manner, and they make some very intelligent conclusions about the system in general. The point out the heterogeneity of the clients connecting in the networks, and how P2P systems need to allow for very low or high latency/bandwidth connections. It's interesting to see how they showed gnutella is vulnerable to an intelligent attack on the server like clients in the system to disintegrate the network into many components. one of their main conclusions is that many of the characteristics of P2P systems match those of classic client-server models. They show the nature of difference between upload and download speeds of the uses. Naturally, the upstream bandwidth will be lower, since a lot of the home users will be on broadband connections that provide caps on upload capacities (to discourage running servers, and also the inherent bias in ADSL connections). I don't think Fig.4 is very relevant at all, since voluntary bandwidth reports from users mean nothing at all, since many users will simply use the default value given or lie. They do mention this in the paper. Figure 5 had an interesting conclusion about the banding of latencies from cross-continent users and those in Europe. All in all,it was a very well-written paper with interesting results. Their methodology was clear and hard to fault. From teifel@csl.cornell.edu Thu Dec 6 12:10:21 2001 Received: from disney.csl.cornell.edu (disney.csl.cornell.edu [132.236.71.87]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6HAK602005 for ; Thu, 6 Dec 2001 12:10:20 -0500 (EST) Received: from localhost (teifel@localhost) by disney.csl.cornell.edu (8.11.3/8.9.2) with ESMTP id fB6HAEH52202 for ; Thu, 6 Dec 2001 12:10:14 -0500 (EST) (envelope-from teifel@disney.csl.cornell.edu) X-Authentication-Warning: disney.csl.cornell.edu: teifel owned process doing -bs Date: Thu, 6 Dec 2001 12:10:14 -0500 (EST) From: "John R. Teifel" To: Subject: 615 PAPER 76 Message-ID: <20011206111728.O51708-100000@disney.csl.cornell.edu> MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII A Measurement Study of Peer-to-Peer File Sharing Systems: This study seeks to characterize the population of end-user hosts that participate in peer-to-peer systems. Important characteristics of end-user hosts that have not been previously analyzed are bottleneck bandwidths between hosts and the Internet, IP-level latencies to send packets to these hosts, how often hosts connect and disconnect from the system, how many files hosts share and download, the degree of cooperation between the hosts, and several correlations between these characteristics. They show that in these peer-to-peer systems (napsters & gnutella) there is significant heterogeneity and a lack of cooperation across peers. They used gnutella and napster crawlers to go over the network and obtain a near instantaneous snapshot of the network properties. Then they measured the network of each system over a period of days. Their napster crawler generated less results because of complaints from users. This study provided extensive and detailed information about actual performance on peer-to-peer systems. They show that many of the peer-to-peer protocols and algorithms currently being proposed are too ideal because they assume all nodes will tend to participate and contribute equally in the system, whereas their measurements showed otherwise. They further suggest that napster and gnutella actually resemble the classic server-client model, rather than a true peer-to-peer model. From c.tavoularis@utoronto.ca Thu Dec 6 12:15:42 2001 Received: from bureau6.utcc.utoronto.ca (bureau6.utcc.utoronto.ca [128.100.132.16]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6HFg602946 for ; Thu, 6 Dec 2001 12:15:42 -0500 (EST) Received: from webmail4.ns.utoronto.ca ([128.100.132.34] EHLO webmail4.ns.utoronto.ca ident: IDENT-NOT-QUERIED [port 37468]) by bureau6.utcc.utoronto.ca with ESMTP id <240027-22887>; Thu, 6 Dec 2001 12:15:31 -0500 Received: by webmail4.ns.utoronto.ca id <164259-209>; Thu, 6 Dec 2001 12:15:21 -0500 To: egs@CS.Cornell.EDU Subject: 615 PAPER 76 Message-ID: <1007658911.3c0fa79fd9530@webmail.utoronto.ca> Date: Thu, 06 Dec 2001 12:15:11 -0500 (EST) From: c.tavoularis@utoronto.ca MIME-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 8bit User-Agent: IMP/PHP IMAP webmail program 2.2.3 This paper provides an in-depth study of two popular peer-to-peer file sharing systems Gnutella and Napster. The authors try to characterize the end-user hosts that make use of these systems in terms of bandwidth, IP-level latencies, connecting and disconnecting patterns, sharing patterns, and cooperation between hosts, in the hopes of making the systems more efficient. Both Gnutella and Napster allow files to be stored on individual users’ computers, and the exchange of files is accomplished through direct uploading or downloading between peers. In Napster, a cluster of central servers maintains a list of files in the system, such that each peer is connected to a central server. Central servers keep track of metadata for each peer including length of connection and reported bandwidth, so that other peers can evaluate their options. In Gnutella, peers form an overlay network, and peers initiate a controlled flood by sending a query to all its neighbors. The query gets propagated, and peers that have a matching file forward it back to the requestor. Ping and pong messages maintain who is in the network, and Gnutella could have disjoint overlay networks. The motivation of this study is to address peers in these systems according to their characteristics, such as to determine how much a load a peer can support. The authors created both Napster and Gnutella Crawlers to gather statistical information and metadata of a large number of users by simply acting like peers in the respective systems. The fact is that peer systems are very heterogeneous and have very different capabilities. A more important result is that it is clear that peers are purposely untruthful about their resources and most often try not to assume the role of a server. Therefore, a small fraction of the peers are actually acting as servers, making the system unbalanced and less efficient. The results presented in this article provide a thorough study of usage in peer- to-peer file sharing systems, and demonstrate the abuse existing in such systems. To provide some ‘lawful enforcement’ rather than a trust-based system, there must be an incentive for peers to be truthful, and/or include direct measurement. As a peer-to-peer user, it is true that I would not gladly leave my peer-to-peer system active when I’m not using it just for the benefit of others. A good question is what could an incentive be? Also, considering this paper was written a couple of years ago when there existed a higher number of low bandwidth users (not to mention dial up connections often meant frequent disconnection, as opposed to cable) it can easily be assumed that currently peer-to-peer systems consist of a healthier balance of clients and servers. From samar@ece.cornell.edu Thu Dec 6 13:09:11 2001 Received: from memphis.ece.cornell.edu (memphis.ece.cornell.edu [128.84.81.8]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6I9B612242 for ; Thu, 6 Dec 2001 13:09:11 -0500 (EST) Received: from aquinas.ee.cornell.edu (aquinas.ee.cornell.edu [128.84.236.57]) by memphis.ece.cornell.edu (8.11.6/8.11.2) with ESMTP id fB6I5bh21445 for ; Thu, 6 Dec 2001 13:05:37 -0500 Date: Thu, 6 Dec 2001 13:07:23 -0500 (EST) From: Prince Samar X-Sender: samar@aquinas.ee.cornell.edu To: egs@CS.Cornell.EDU Subject: 615 PAPER 76 Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII 76) A Measurement Study of Peer-to-peer File Sharing Systems This paper presents an interesting study of two of the most popular peer-to-peer systems, Napster and Gnutella. The authors made various measurements of these systems to study the bottleneck bandwidth, latency, availability and file sharing patterns of the peers in the network. Based on this study, the authors make a number of recommendations for effectively designing a good peer-to-peer system. One major difference between Napster and Gnutella is that Napster relies on a cluster of centralized servers to maintain an index of files that are currently being shared by active peers. Gnutella does not rely on any central servers. Instead, it forms an overlay network using ping and pong messages. A peer initiates a controlled flood of the network to search a file. The authors designed crawlers for the Napster and the Gnutella systems to make the measurements. Their study shows that the assumption that all nodes are identical and should be treated equally in a p2p system is not correct. With respect to a number of characteristics like uplink/downlink bandwidths, latencies, lifetimes, shared data etc, the peers greatly differ from each other. Also, many peers are found to report incorrect parameters or not report at all. Also, some peers are found to be free-riders on th network, just downloading files but not offering them for others. In fact, the experiment data indicates that in many ways, the Napster and Gnutella systems resemble the classic client-server model. Based on these results, the authors recommend that future p2p systems should delegate different degree of responsibility to different nodes depending on their particular characteristics. The systems should be robust, being able to measure these characteristics and not relying on what is actually reported. Some inbuilt incentive should be there in the systems for peers to cooperate with others, offering its own resources. The system should be able to adapt to the differences in the host and network characteristics. From papadp@ece.cornell.edu Thu Dec 6 13:16:39 2001 Received: from memphis.ece.cornell.edu (memphis.ece.cornell.edu [128.84.81.8]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6IGd613342 for ; Thu, 6 Dec 2001 13:16:39 -0500 (EST) Received: from kiki.ece.cornell.edu (kiki.ece.cornell.edu [128.84.83.13]) by memphis.ece.cornell.edu (8.11.6/8.11.2) with ESMTP id fB6ID5h21707; Thu, 6 Dec 2001 13:13:05 -0500 Date: Thu, 6 Dec 2001 13:19:58 -0500 (EST) From: "Panagiotis (Panos) Papadimitratos" To: Emin Gun Sirer cc: Panagiotis Papadimitratos Subject: 615 PAPER 76 Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Review of: 'A Measurememnt Study of Peer-to-Peer File Sharing Systems,' by S. Saroiu, P.K. Gummadi, S.D. Gribble The authors present measurements for two peer-to-peer systems in order to identify the characteristics of the participating processes and their interaction with the underlying network. Then, they provide a set of design guidelines based on their experimental results. Their tools 'discovered' a population of nodes implementing the peer processes and then measurements of latency, lifetime, bottleneck bandwidth (up/down stream), node uptime, session duration, number of shared files, up/downloads were measured. Obviously, a dependence on the last-hop BW is observed, since end-users are far from giga-bit trunks. Latencies are correlated to bottleneck BW but on one hand they either cease to drop although BW increases, or remain low despite a relatively lower bw, if the core network is fast. Of course, the traffic is exchanged at the application layer, and locality and caching are not there at all. Such measurements are also affected by the actual sought content and the advertised, by the acting-as-server end process, resources. On one hand, this depends on the willingness/incentive of users to share content and allocate resources. This is the reason for the authors to propose the design of a system that does not accept parameter values from the peer process but self-configures based only on measurements. Nevertheless, measurements of good value would have to be performed over a long enough period and a realatively stable network state, which makes this approach less attractive, to say the least, for 'real-time' (or real-world) application. From andre@CS.Cornell.EDU Thu Dec 6 14:16:36 2001 Received: from postoffice.mail.cornell.edu (postoffice.mail.cornell.edu [132.236.56.7]) by sundial.cs.cornell.edu (8.11.3/8.11.3/M-3.7) with ESMTP id fB6JGa624884; Thu, 6 Dec 2001 14:16:36 -0500 (EST) Received: from khaffy (d7b106.dialup.cornell.edu [128.253.157.106]) by postoffice.mail.cornell.edu (8.9.3/8.9.3) with ESMTP id OAA13443; Thu, 6 Dec 2001 14:16:34 -0500 (EST) Received: from andre by khaffy with local (Exim 3.33 #1 (Debian)) id 16ByPK-0000xw-00; Thu, 06 Dec 2001 14:17:46 +0100 Date: Thu, 6 Dec 2001 14:17:46 +0100 From: =?iso-8859-1?Q?Andr=E9?= Allavena To: =?iso-8859-1?B?R/xu?= Sirer Cc: andre@CS.Cornell.EDU Subject: 615 PAPER 76 Message-ID: <20011206131746.GA3706@khaffy> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Disposition: inline Content-Transfer-Encoding: 8bit User-Agent: Mutt/1.3.24i Sender: =?iso-8859-1?Q?Andr=E9_Allavena?= A Measurement Study of Peer-to-Peer File Sharing System This papers presents a detailed study of the hosts constituing Gnutella and Napster (May 2001) The hosts are quite different, with different bandwidth, and different behaviours. Some have the profile of servers, some have the profile of clients (and those are a significant portion: 25%). Napster is predominant on hosts with very slow bandwidth link (Popularty of Gnutella among geeks? - ressources consumptions?) Also a good number of hosts underestimated their actual bandwidth, because it means less people are going to download from them. Napster was vulnerable to the shutdown of their centralised servers, but Gnutella also is quite vulnerable if you should down the right servers. This paper remainds us that there have to be incentives for people to cooperate, trust is far from enough. This opens the question of how to develop shared systems (such as p2p) where trust shouldn't be granted but should be enforced by the protocols. I really like that paper btw. -- André Allavena (local) 154 A Valentine Place École Centrale Paris (France) Ithaca NY 14850 USA Cornell University (NY) (permanent) 879 Route de Beausoleil PhD in Computer Science 06320 La Turbie FRANCE