Keywords: Peer-to-Peer Networks; Free Riding; Connection Management; Distributed Systems

Transcription

1 Elsevier Editorial System(tm) for Computer Communications Manuscript Draft Manuscript Number: COMCOM-D R1 Title: A Connection Management Protocol for Promoting Cooperation in Peer-to-Peer Networks Article Type: SI Foundations of Peer to Peer Computing Keywords: Peer-to-Peer Networks; Free Riding; Connection Management; Distributed Systems Corresponding Author: Prof. Ibrahim Korpeoglu, PhD Corresponding Author's Institution: Bilkent University First Author: Murat Karakaya, PhD Candidate Order of Authors: Murat Karakaya, PhD Candidate; Ibrahim Korpeoglu, PhD; Ozgur Ulusoy, Ph.D. Abstract: The existence of a high degree of free riding in Peer-to-Peer (P2P) networks is an important threat that should be addressed while designing P2P protocols. In this paper we propose a connection-based solution that will help to reduce the free riding effects on a P2P network and discourage free riding. Our solution includes a novel P2P connection type and an adaptive connection management protocol that dynamically establishes and adapts a P2P network topology considering the contributions of peers. The aim of the protocol is to bring contributing peers closer to each other on the adapted topology and to push the free riders away from the contributors. In this way contribution is promoted and free riding is discouraged. Unlike some other proposals against free riding, our solution does not require any permanent identification of peers or a security infrastructure for maintaining a global reputation system. It is shown through simulation experiments that there is a significant improvement in performance for contributing peers in a network that applies our protocol.

2 Cover Letter April 17, 2007 Dear Editor: We revised our paper and we are submitting the revised version. We are also submitting a document describing how we addressed the reviewers comments. Please let us know if you need anything else. Best Regards Ibrahim Korpeoglu

3 * Detailed Response to Reviewers Response to Reviewers Comments Paper Title: A Connection Management Protocol for Promoting Cooperation in Peer-to- Peer Networks First of all, we would like to thank reviewers for their constructive and careful comments, which helped us very much to improve the quality of our work. To address the concerns of reviewers, we conducted extensive new simulation experiments, and modified the content and presentation of the paper significantly. In summary, in this revision: We re-conducted several simulation experiments to observe the effects of the number of peers, and the number of Free Riders by increasing the scales of these parameters, and presented the new results. In general, our protocol PCMP was observed to scale well due to its distributed nature. In response to the reviewers comments, we relaxed and extended several assumptions (file size, file popularity, etc.) and conducted new simulation experiments. These new results showed that PCMP adapts well to the varying values of different parameters of P2P environment. We carefully examined and worked on the concerns raised by the reviewers. We updated the context and presentation of the paper accordingly. Furthermore, we developed the paper by adding new sections, which discuss some important issues raised by the reviewers. We added a new replacement algorithm (Sized-Based PCMP) thanks to reviewers comments. We updated the paper with new references recommended by the reviewers. We fixed the typos as suggested by reviewers. In the following we provide our response to each reviewer s comments. 1

4 Reviewer #1: 1) Writing is generally good, but I think it still needs some revisions. For example, at page 3, the first paragraph of Related Work section mention "73% of them share ten or fewer files" while them refers to "70% of the peers do not share any files at all" and these two contradict. Thank you. The typo was fixed. Now, it reads: It was reported that 66% of the peers do not share any files at all, while 73% of them share ten or fewer files. 2) The authors do not provide their idea with analytical discussion. Additionally, there is even no section to clarify the "rationale behind the idea". It is very difficult to understand why the proposed protocol works well. To address the reviewer s comment, we updated the third section (P2P Connection Management Protocol) and added the Our Approach and Motivation section with an analytical discussion. 3) Why the percent of free riders is fixed at 70% (Table 1). It is expected the authors to verify the operation of their proposed protocol under different population of free-riders. For example, the protocol should work well where the number of Free Riders is negligible. To answer the reviewer s comment we extended the scale of the simulation tests in Section to cover the free rider population from very low (30%) to very high (90%). As shown in (the new version of) Figure 18, even with a low population of Free Riders, the protocol continues to perform well. We updated Section and Figure 18, accordingly. 4) The number of copies for each file is identical and fixed to four (P-16). It is far from the real world scenarios which some files are rare and some other plentiful. This could change the topology of the network considerably. To address the reviewer s concern we set up new simulation tests with three different file distributions, called RARE, POPULAR, and UNIFORM. With these distributions, we considered different levels of file replication. We observed through performance experiments that with different replications schemes tested, our protocol PCMP performs very well. Details of the performance results were added to the paper as a new section (4.3.5 Effects of Different File Sizes and Popularity). 2

5 5) "We assumed that each file was of the same size". This assumption could not acceptable. Because the size of the shared files could considerably affect the consumed bandwidth of the container, which is a major incentive for free-riders. To address this comment, we added a new section (4.3.5 Effects of Different File Sizes and Popularity). In this section, we relaxed the file size assumption and provided a new replacement algorithm (S-PCMP) for PCMP to clarify how PCMP can handle the requests for different file sizes. New replacement algorithm takes file sizes into account by recording each neighbor s total uploaded file size when selecting a victim neighbor. Thus, only uploading small size files would not help a free rider to walk around the protocol. 6) Numerous assumptions are not justified. For example, they mention that "We assumed TTL is set to be 3 hops." So why? As we stated in Section 4.1 (Overview of the Simulation Model), the parameter values were chosen so as to be comparable to the related works such as [1,2,6,32,33] 1. Those works observed the actual P2P network traffic and reported its characteristics. For example, in [1], Adar and Huberman state that specifically, they found that nearly 70% of Gnutella users share no files. Based on that observation, we set the default Free Riding population ratio to 70% and their shared file ratio to 1%. Furthermore, the assumptions we adapted in our simulation model are similar to those used in other related works such as [14,16,17,18] 2. For the TTL value, Gnutella Protocol v0.4 leaves TTL field value unsigned. In real life applications, TTL is usually set to 7. We set it to 3 in our simulation tests, since the network topology we simulate is small compared to the real world. If we had set TTL to 7, then most of the queries would have covered almost all of the peers, which would not have been realistic. Therefore we chose 3 as a reasonable value for TTL. Also, we observed that changing the TTL value does not have an impact on the relative performance of Gnutella and our PCMP protocol. These facts were added to the paper in Section ) The achieved results reported in Fig. 11 are not considerable. Isolating just "a total of 24 free riders (out of 630)" is not sufficient to name a protocol successful. This fact is hold also for Fig. 12. In Section 4.3.1, we discussed the observed effects of PCMP over P2P network topology for 3 important metrics: 1- The number of connections established among contributing peers, 2- The number of OUT-connections of free riders to contributing peers, and 3- The number of isolated free riders. Considering these 3 criteria together, we can appreciate the success of the protocol. With the current settings of the simulation, we have observed that the connectivity among contributors was increased by 82%, the number of OUTconnections from free riders to contributors was decreased about 67%. In Fig 11, the number of totally isolated free riders seems to be fewer but it is only one of the effects aimed by the protocol. Actually, if we extend the simulation time we would observe more FR to be isolated. 1 All the referenced papers are given from the reference list of the paper. 2 All the referenced papers are given from the reference list of the paper. 3

6 Similarly, the results provided in Fig 12 showed that the free rider downloads were dropped by 15%. The other performance results showed that the protocol was also successful according to all defined criteria with higher level of improvement. For example, the contributor downloads were increased by 50%, the download cost of contributors were decreased about 30%, and the network traffic was reduced 36%. Considering the results for all the performance metrics, we believe that PCMP is successful. 8) The authors failed to compare their results with the previous protocols. It is unclear that does the proposed protocol outperform the others or not? We compare our protocol PCMP and its variations with one of the most common unstructured P2P protocols, namely Gnutella. Gnutella protocol is the only available open P2P protocol. The other popular protocols, such as edonkey 2000 and FastTrack, have not been publicized. Therefore, we preferred to compare PCMP with the well-defined Gnutella protocol. We implemented almost every aspect of Gnutella in the simulation and executed fairly realistic simulation test. There exist some other protocols proposed against FR. However, none of the papers which present those protocols include implementation information detailed enough for an extensive comparison. In the following, we provide the results of a rough comparison of our approach with an incentive model, called SLIC, which was proposed by Sun and Garcia-Molina for promoting cooperation in unstructured P2P networks [14] (see our Related Work section). Further analysis of the comparative performance is left as a future work due to the time constraint for the revision. Similar to our proposal, the SLIC method is also based on the local interaction of peers and management of connections to encourage cooperation. Each peer assigns weights to its neighbors based on the behavior of the neighbors, and those weights determine the amount of connection capacity assigned to the neighbors. To assign connection capacities, each peer counts the number of query hits it receives via its neighbors and updates weights based on these numbers. The weights moderate the future query processing capacity over that connection between these peers. Their work has some similarities with our mechanisms. Here, we provide the results of a comparison of SLIC and our protocol PCMP. The results were obtained in our simulation environment in which we implemented an adapted version of SLIC as well. Experimental results are summarized in the following Table. Metric T-PCMP SLIC Gnutella # Downloads of FRs 2321 (-16%) 2581 (-6%) 2763 # Downloads of Contributors 1772 (+52%) 1095 (-6%) 1172 Comparative Performance Results against SLIC In the original SLIC paper [14], the simulations are performed only for the scenarios where only a single probe node is selected in a P2P network to act selfishly and as a free rider. Consequently, it is not clear how that model would react when most of the peers would become free riders rather than just a small number [29]. In our simulation experiments, however, we obtained SLIC results for also the scenarios where there exists a prevalent free riding in the simulated P2P network and where all peers apply the SLIC mechanism. The first important result that we obtained in our simulation environment about the SLIC is that, SLIC could not differentiate or correctly identify contributors and free riders. Therefore, 4

7 it causes a decrease in the number of downloads of not only free riders, but also contributors. In our experiments, the amount of downloads that can be performed by peers (which could be either free riders or contributors) is decreased by about 6%. Our scheme, however, does not cause a reduction in the contributors downloads. There can be several reasons why SLIC can not differentiate well between free riders and contributors. One of these reasons is the fact that each node applying SLIC monitors the number of all query hits it receives via its neighbors and updates weights which moderate the future query processing capacity it offers to others. That is, SLIC evaluates the contribution of the subnetwork reachable via each neighbor to decide on the amount of service to be provided to these neighbors. In our proposed framework, however, we assess the contribution of each individual peer to any other peer. In other words, in SLIC, peers are not only responsible for their performance, but also for the performance and amount of contributions of their neighbors. Considering the prevalence of free riders, even the contributors cannot provide a better service than the services provided by all the peers connected through them. Because, even a peer increases its contribution, it cannot have any control over the other peers to increase their contributions. At the end, in SLIC, the monitoring peers punish or reward the controlled peers not only according to their contribution but also according to the contribution of all the peers connected through them. We executed several experiments to observe the effect of increased individual contribution on the performance of a single peer. We selected a peer randomly, and increased its shared file size in each experiment. However, we could not observe any meaningful increase in the service it gets. This observation is in line with the above arguments. [14] Qixiang Sun and Hector Garcia-Molina, SLIC: A Selfish Link-based Incentive Mechanism for Unstructured Peer-to-Peer Networks, Proceedings of the 24th International Conference on Distributed Computing Systems (ICDCS 2004), [29] D. Hales and B. Edmonds, Applying a socially-inspired technique (tags) to improve cooperation in P2P Networks, IEEE Transactions in Systems, Man and Cybernetics - Part A: Systems and Humans,Volume 35(3), pp , ) It seems that by the proposed protocol, it takes some considerable times for a new visitor of the network to have a considerable usage of the network. This is a key idea in encouraging the new users and extending the network. When new visitors arrive, they first need to solve a proof-of-work (POW). After that, they can connect to some other peers as the Gnutella Protocol states. In PCMP, new comers have a limited number (4 in the current simulation settings) of IN and OUT connections. After successfully connected to network, they can begin to search using their existing OUT connections. However, if they are Free Riders they will loose their OUT-connections in course of time. Thus the chance of a hit and consequently a download will be decreased. Actually, in the paper we had presented the related results to support this argument in Section (Reactiveness of PCMP). To understand how much time it takes for a new visitor to begin to use the network, we probed 20 contributors and 20 free riders and recorded their first three downloads time. As seen the below figure, the first download of free riders and contributors was executed around 600 simulation unit of time. The third download time is 1315 for contributors and 1566 for free riders. Please note that in the simulation, a file is assumed to be downloaded in 60 simulation units of time, and after each download or a search, peers sleep for 60 simulation units of time to search again. Thus, it does not take much time for a new visitor to begin use of the network. 5

8 In Figure 2, we marked the 20 peers download times individually. As seen in the figure, regardless of their contribution, almost all the peers downloaded the first file after a similar waiting time. However, when PCMP observed their contribution levels, the number of downloads and the second and third download times began to delay. Free riders were negatively affected due to the effective strategy of PCMP against free riding. Figure 2 Downloads of the Contributors (on the left) and Free Riders (on the right) Below, we summarize the number of downloads. As seen in the table, PCMP differentiates the service level (downloads) to a new comer depending on its contribution; even all the new comers begin with the same number of connections. Number of Contributors Number of Free Riders 1 st Download nd Download rd Download As a result, we conclude that the protocol does not cause a considerable waiting time for the new comers as long as they solve the POW and they are not Free riders. These observations are actually a motivation for peers to cooperate. If they share, they would receive more downloads in less time. If they do not share, they cannot download more and they need to wait for longer time. 6

9 10) The embedded captions of figures 17 and 18 are not readable. Fixed: All the captions in the figures are now readable. Thank you. 11) In related work section, it would be nice to mention the following paper which is based on the topology control: Y Chawathe, S Ratnasamy, L Breslau, N Lanham, and S Shenker. Making Gnutella-like p2p systems scalable. In Proceedings of the 2003 conference on Applications, Technologies, Architectures, and Protocols for Computer Communications, pages , Karlsruhe, Germany, The Related work section now includes a brief discussion of the above work. 7

10 Reviewer #2 1) One criticism of the algorithms used to manage the IN and OUT connections is that the number of downloads is considered and not the size of each download. A free rider can filter out the requests that are lower than a certain size and only serve the smaller downloads, hence obtaining credit for getting away with rejecting large requests. It is not clear how does the size of the download requests play out. This issue needs to be clarified, as the main concern for free riders is bandwidth utilization, not the number of downloads accepted. It is observed that most of the peers in P2P networks do not share any files at all [1,2,6,19,20] 3. For example, in [1], Adar and Huberman concluded that Specifically, we found that nearly 70% of Gnutella users share no files... Thus, the main problem is the fact that most of the peers do not share small or big size files at all. If we can enforce or motivate them share some files small or big- it would alleviate the current situation. If a peer searches for a file small or big size- and finds it at a peer, then its request is satisfied. Thus, in a peer s point of view, the size of the downloaded file is not important. What is important for him is to find it or not. If the free riders begin to share small size files to save their bandwidth and if they receive requests to these files and upload them to the requesters, there is nothing wrong with that. If they provide small size files, which are not requested by anyone, then they cannot benefit from that. PCMP rewards the provider with an OUTconnection only when a download occurs. To improve the work and to address the comment, we added a new section (4.3.5 Effects of Different File Sizes and Popularity). In that section, we relaxed the identical file size assumption and provided a new replacement algorithm (S-PCMP) for PCMP to clarify how PCMP can handle the requests for different file sizes. The new replacement algorithm takes file sizes into account by recording each neighbor s total uploaded file size when selecting a victim neighbor. Thus, uploading only small size files would not help a free rider to walk around the protocol. 2) In addition, the algorithms are heuristic, and the approach is not sufficiently motivated. It seems that the overhead is large in this case, and needs all the existing P2P network topology formation algorithms to be changed drastically, unlike a solution such as AOTO which could be added to improve topologies and search efficiencies without changes to existing widespread algorithms. Making this approach similar in essence to AOTO could help its validity and acceptability. In order to explain our heuristic and the motivation behind it, we extended the third section and added a new subsection (3.1 Our Approach and Motivation). -Overhead: In PCMP, peers have to manage two neighbor lists: IN and OUT connected neighbors. The data structures of these lists, presented in the paper, are very simple tables with three fields. Due to the power-law distribution of node degrees observed in P2P networks [4,34] 3, we expect the average number of neighbors of a peer to be around 3-4, and therefore the space overhead imposed on each peer will not be large. The computation overhead of updating these lists would not be large either. As we stated in the paper, to reduce the computation overhead, these lists can be updated periodically rather than with every upload/download operation. In summary, the space and computation overhead of the PCMP is not large. 3 All the referenced papers are given from the reference list of the paper. 8

11 -Implementation of PCMP: As the reviewer point out, we can relate the PCMP implementation with AOTO. Similar to AOTO s proposed implementation approach over Gnutella, PCMP can be implemented by modifying only the client software, not the protocol messages. Because, the Gnutella message types and formats would remain the same. However, the client software has to be modified to recognize PCMP connection types and management of them. In AOTO, it is stated that Instead of flooding to all neighbors, SF uses a more efficient flooding strategy to selectively flood a query on an overlay multicast tree., and each peer has a neighbor list which is further divided into flooding neighbors and non-flooding neighbors in SF.. These facts are very similar to PCMP implementation requirements. In PCMP, we do not flood the queries to all neighbors either, we only route queries to OUT-connected peers. Again, as in AOTO, each peer has two neighbor lists: IN-connected and OUT-connected. Thus, the implementation of AOTO and PCMP into a Gnutella like unstructured P2P network is not different from each other, and is not difficult. In essence, similar to AOTO, the implementation of PCMP does not require drastic changes in the P2P protocol. 3) A criticism of the simulations performed is that it doesn't look at the size of the downloads, as mentioned before. This comment is addressed in the first answer to the reviewer. 9

12 Reviewer #3: 1) The paper starts well. It's readable and interesting. The related work section is fairly detailed (though see below). The concept is explained quite well in section 3, although and were a bit confusing. I would also prefer to see the overview from 3.4 somewhere before going into the detail of the algorithms. We rewrote the parts of Sections and to remove confusion. As suggested by the reviewer, we added an overview of Section 3.4 into Section 3.1 by modifying the paragraph before the last. 2) The evaluation is a bit weak with so few peers. In Section (Effects of Peer and Free Rider Population), we reported the results about the effects of peer population size on the performance metrics. To obtain those results, we performed experiments with various settings of the peer number parameter: 100, 400, 900, and 2500 peers. We observed that the relative performance results we obtained with these different settings of P2P network size are similar for the Gnutella and PCMP protocols. This is due to the fact that PCMP is based on interactions between a pair of peers. Due to the fifth comment of the reviewer, we extended these simulation tests to cover 4900 peers. (Please see the related discussion in response to the fifth comment of the reviewer.) The new results still support the scalability of the protocol. Thus, the results obtained by simulating 900 peers are in parallel with the results obtained by higher peer population. Further increase in the number of peers did not have an effect on the relative performance, while causing a drastic increase in the time required performing the simulation experiments. Therefore, we restricted the increase in the number of simulated peers to ) Is the error too small to include in graphs? If so, say this, otherwise include it. Yes, the error is too small. 95% confidence intervals were obtained for the performance results. The width of the confidence interval of each statistical data point was observed to be very small. As suggested by the reviewer we included this fact in Section ) Graphs are difficult to read (text should be no smaller than caption). Fixed. All the captions in the figures are now readable. 5) says the protocol should scale well, although experimental evidence only progresses to 2500 peers. It's surprising how many unexpected effects occur when networks become larger. In particular, I wonder if there will be some visible effect at larger sizes caused by the limited number of connections each peer can maintain. To address the reviewer s concern, we extended the scale test to 4900 peers. From the results of re-executions, we observed that with the increased number of peers, PCMP still performs very well. This is an expected result, since we use a distributed approach, which is scalable. As explained in the paper, PCMP is based on the interactions between each pair of peers. Therefore, we can safely argue that the connection management mechanism is not directly dependent on the number of peers. 10

13 As seen in the figure, it can be safely argued that the proposed protocol is not negatively affected from increasing peer population. PCMP provides similar level of benefits no matter what the number of peers is. Another concern of the reviewer is the number of a peer s connections as the peer population increases. In [4,34] 4, it was observed that distribution of node degrees follows a power-law distribution in real life Gnutella network. In [4], the average number of neighbors of a peer is expected to be around 4. Similarly, in [34] the average node degree is reported as 5.5. In essence, in our context, we do not see any problem in implementing our protocol in real life P2P networks since the average number of neighbors of a peer stays fairly low in spite of the great peer population. These observations are summarized in Section ) I'm glad attacks were discussed but more is needed here, especially with respect to whitewashing (peers rejoining the network with new identities) in 5.3. This is absolutely fundamental to the protocol - if it cannot be mitigated, there is no point implementing PCMP. POW has been shown not to work for ("Proof-of-work proves not to work" by Laurie and Clayton, 2004). Are there problems getting it to work with P2P (zombie peers)? If not, this should at least be addressed. Also, could malicious peers pose as multiple peers simultaneously, each with their own fresh set of OUT-connections? In the referenced work [31] 4, the authors estimated that each user received, on the average, about 50 legitimate and about 60 spam s per day. They calculated that if spammers use their own machines and if each calculation of POW takes 50 seconds, then spammers could send maximum 1750 mails per day instead of mails per day. In spite of this considerable reduction, they argued that POW would have a side effect on legitimate users:... (real users) would be unable to provide proof-of-work for all the legitimate ..., making the system completely infeasible to deploy. Because they estimated that..., at the values calculated above of 1,750 s per day, a proof-of-work scheme would prevent legitimate activity by 0.13%... All the estimations in [31] are based on the goal of reducing spam to 1% of normal s received. Using their numbers, we can restate this goal as receiving only 0.5 spam per day instead of 60 spams per day. If this goal is changed to 2% (1 spam per day), the side effect mentioned above would not occur at all. Because then we would change the POW duration accordingly and let the number of mails (spam or normal) per day be Thus, this new threshold would not limit the legitimate users activities and the system would be feasible to be 4 All the referenced papers are given from the reference list of the paper. 11

14 implemented. In essence, POW can be used to considerably reduce the number of spam mails effectively and efficiently. However, it is not possible to achieve much lower levels of this reduction. But still, the possible outcome would be very satisfactory almost for all users. In our context, we believe that we can implement POW as an effective discouraging method against Free Riders. There are no side effects similar to the ones mentioned above in our application. Sure, to prevent and avoid whitewashing totally is not possible for the time being in unstructured P2P networks. On the other hand, using POW we can filter and restrict these attacks to very low levels. We added the referenced paper [31] and the above discussion into Section 5.3. [31] B. Laurie and R. Clayton, Proof-of-work proves not to work, The Third Annual Workshop on Economics and Information Security, Also, could malicious peers pose as multiple peers simultaneously, each with their own fresh set of OUTconnections? Free riders can use multiple peers to connect to P2P network, as it is possible in any P2P systems. However, this attack can be reduced to the whitewashing attack discussed in the paper (Section 5). Also, since each peer is required to solve a POW, the multiple copies of the peer will have to spend a considerable effort to connect to the network. 7) Related work: in section 2, it's claimed that the idea of rearranging the topology to disadvantage free-riders is novel. This is not the case of course. The novelty is the idea of "one-way connections", but shuffling free-riders to the outskirts of unstructured networks has been explored indirectly with reputation and incentive schemes: "Foreseer: a novel, locality-aware peer-to-peer system architecture for keyword searches" by Cai and Wang (2004). "Peer-to-peer's most wanted: Malicious peers" by Mekouar et al (2005). "Incentive mechanism for peer-to-peer media streaming" by Habib and Chuang (2004). and more similarly, through emergent structure in SLAC: "From selfish nodes to cooperative networks - emergent link-based incentives in peer-to-peer networks" by Hales (2004). We agree with the reviewer s comment. In the Abstract we emphasize what is novel in our work: In this paper we propose a connection-time solution that will help to reduce the free riding effects on a P2P network and discourage free riding. Our solution includes a novel P2P connection type and an adaptive connection management protocol that dynamically establishes and adapts a P2P network topology considering the contributions of peers. Similarly in the Related Work section, we state that: Our proposal in this paper, the P2P Connection Management Protocol (PCMP), is another solution to the free riding problem with an approach that is quite different than the methods mentioned above. The methods mentioned before in the text are incentive-based and reciprocity-based schemes. After explaining our proposal, we continue that: There exist some other studies which also focus on modifying P2P topology such as [23 25]... In summary, as the reviewer suggests, we now state that our proposal is not the first work aiming to push Free Riders from P2P network by changing the topology. 12

15 Instead, we state that the connection types and the management strategy of these connections proposed to prevent free riding are novel. The reference [Cai and Wang, 2006] has been added to the Related Work section. 8) Typos: Page 3: "They also found a high level of..." (remove "out"). Page 6: "...will have difficulty in reaching the resources..." (remove "to"). Fixed. Thank you. 13

16 EDITOR'S COMMENTS 5. This is one of good papers in the category FAIR RESOURCE SHARING sub-area. This is a strong contender for second round. 6. Highlight your principal innovation/contribution in more precise terms (at the level of P2P researcher) in your abstract and introduction. Particularly highlight the exact problem you have solved and the exact analytic contribution of your work (the looking for work with formal/analytical basis). In the new version of the paper, innovation and contributions of our work are clearly stated as well as the exact problem we deal with and the exact analytical contribution of our work (please see the Abstract, Introduction, and the Our Approach and Motivation section (Section 3.1)). 7. Include reference to closest conference publication (if any). It has to be substantially different (even textually) to be accepted. Include comparison in text and send us reference publication as attachment. The work presented in this paper has not been submitted to any conference. 8. Address all concerns of the reviewers. We have addressed all the concerns of the reviewers. 14

17 * Manuscript A Connection Management Protocol for Promoting Cooperation in Peer-to-Peer Networks Murat Karakaya, İbrahim Körpeoğlu, Özgür Ulusoy Department of Computer Engineering, Bilkent University, 06800, Ankara, Turkey Abstract The existence of a high degree of free riding in Peer-to-Peer (P2P) networks is an important threat that should be addressed while designing P2P protocols. In this paper we propose a connection-based solution that will help to reduce the free riding effects on a P2P network and discourage free riding. Our solution includes a novel P2P connection type and an adaptive connection management protocol that dynamically establishes and adapts a P2P network topology considering the contributions of peers. The aim of the protocol is to bring contributing peers closer to each other on the adapted topology and to push the free riders away from the contributors. In this way contribution is promoted and free riding is discouraged. Unlike some other proposals against free riding, our solution does not require any permanent identification of peers or a security infrastructure for maintaining a global reputation system. It is shown through simulation experiments that there is a significant improvement in performance for contributing peers in a network that applies our protocol. 1 Introduction Free riding is an important threat against efficient operation of Peer-to-Peer (P2P) networks. In a free-riding environment, a small number of contributing peers serve a large number of peers; many download requests are directed towards a few sharing peers. This situation may lead to scalability problems [3] 1 This work is partially supported by The Scientific and Technical Research Council of Turkey (TUBITAK) with grant numbers EEEAG-104E028, and EEEAG- 105E065. Preprint submitted to Elsevier Science 17 April 2007

18 and to a more client-server-like paradigm [5,6], which overweigh the benefits of P2P network architecture. Additionally, renewal or presentation of interesting content may decrease in time, and the number of shared files may grow very slowly. The quality of the search process may degrade due to an increasing number of free riders on the search horizon. Moreover, the large number of free riders and their queries generate an extensive amount of P2P network traffic, which may lead to degradation of P2P services and inefficient use of the resources of the underlying network infrastructure. There are various reasons for free riding. Bandwidth limitation of peers connections may be one reason. Another reason might be peers concern about sharing bad or illegal data from their own computers, even though they are not concerned about using this type of data. Some peers may also have security concerns when they share resources. In this paper, we propose a connection-based solution against free riding that will alleviate the problems associated with free riding. Our solution involves the definition and use of two new connection types (IN and OUT connections) and a P2P Connection Management Protocol (PCMP) that dynamically establishes the connections between peers, and adaptively modifies the P2P topology in reaction to the contributions of peers. Our protocol promotes cooperation among peers and discourages free riding, and can be used in unstructured P2P networks such as Gnutella [10]. Our claim is that if we can adjust the P2P network topology dynamically in reaction to peers contributions, the adapted topology can favor the contributing peers in getting service from the P2P network. The adapted topology can also exclude the free riders from the P2P network and therefore the adverse effects of free riding can be reduced as well. Furthermore, we expect that our approach will help a P2P network to become more scalable and robust. We did extensive simulations to evaluate our protocol and we have seen significant improvement in the performance of a P2P network with free riders when our solution is applied. The organization of the paper is as follows. In Section 2, we discuss the related work. In Section 3, we describe our solution and the PCMP connection management protocol. In Section 4, we present our simulation model and provide the simulation results. In Section 5, we discuss some possible attacks to our scheme and how we can cope with them. Finally, in Section 6, we give our conclusions. 2 Related Work User traffic on the Gnutella network was extensively analyzed by Adar and Huberman in [1], and it was reported that 66% of the peers do not share any 2

19 files at all, while 73% of them share ten or fewer files. Furthermore, 63% of the peers who share some files do not get any queries for these files; and 25% of all peers provide 99% of all query hits in the network. Saroiu et al. confirm that there is a lot of free riding in Gnutella as well as in Napster [6]. They observed that 7% of the peers provide more files than all of the other peers combined. In a recent work [19] Hughes et al. pointed to an increasing downgrade in the network s overall performance due to free riding. Their results indicated an increasing level of free riding compared to Adar and Huberman s work. For example, they observed that 85 percent of peers share no files at all. They concluded that free riding was becoming more prevalent. In another work, Yang et al. reported their findings about free riding behavior in the Maze P2P system [20]. They also found a high level of free riding, with about 80% of the peers behaving like clients. They observed that client-like users (free riders) were responsible for 51% of downloads, but for only 7.5% of uploads. These statistics suggest the existence of free riding in spite of the incentive mechanism provided by the Maze P2P system. All these observations have caused researchers to be concerned about the free riding problem and to propose solutions. In fact, some mechanisms against free-riding have already been implemented ([20 23]). There are also a number of solutions that have been proposed in research studies ([3,7,8,11,14,24,25]). Existing mechanisms and proposed solutions for the free-riding problem can be categorized into two main groups: a) incentive-based and b) reciprocity-based schemes. Incentive-based solutions have been proposed to encourage user cooperation within P2P systems. One of the most common way of implementing incentives is to apply telecommunications models for pricing network resources by incorporating micro-payments in P2P networks, such as KARMA [8],ARA [24], PPAY [26], etc. In these systems, each user has to purchase service on demand, using a virtual currency that is obtained as payment for providing service in turn. Some other incentive-based approaches implement reputation mechanisms [25,27,28]. Reputation-based approaches depend on identifying and monitoring peers contributions to other peers, and then refusing service to peers with bad reputations. The schemes that depend on micro payments have limitations when applied to many common P2P network architectures. In general, incentive schemes based on persistent identifiers are complicated by the anonymity of peers, by collections of widely dispersed peers, and by the ease with which peers can modify their online identity [7,12]. 3

20 Reciprocity-based schemes have been proposed as non-monetary mechanisms based on reciprocity among peers, such as [3,11,14]. Peers maintain histories of past behavior of other peers and use this information in their decision making processes. These schemes can be based on direct reciprocity (Tit-for- Tat) or indirect reciprocity (Utility-Based). In direct reciprocity schemes, peer A decides how to serve peer B based solely on the service that B has provided to A in the past. In contrast, in indirect reciprocity schemes, the decision of A also depends on the service that B has provided to other peers in the system. However, there are some ways of getting around the utility values. For example, a user can share some small files with fake names resembling popular file names. If other users download these files, that user s utility value will increase. Additionally, relying on information about a peer that is stored and provided by the peer itself may cause problems as well [6]. In [14], the authors propose an incentive model to encourage cooperation in unstructured P2P networks. This model, called SLIC, depends on the local interactions of peers. In SLIC, each peer assigns weights to its neighbors and updates these weights based on the number of query hits it receives via each neighbor. Those weights determine the amount of messaging capacity assigned to each neighbor. In a previous work [11], we also proposed a framework which focuses on detection of neighbors that are free riders and taking counter actions against them. The proposed framework counts both query hits and query messages, and considers the originator and receiver of these messages. Based on this information, peers make a decision about their neighbors. The proposed framework also categorizes the free riders into several categories. This enables the framework to apply several different counter-actions that are tailored to different types of free riding. The framework assesses the contribution of each neighbor both to the monitoring peer and to the overall system. Our proposal in this paper, the P2P Connection Management Protocol (PCMP), is another solution to the free riding problem with an approach that is quite different than the methods mentioned above. The PCMP protocol is based on managing connections among peers to discourage free riding and to provide incentives for cooperation. The scheme is distributed and does not require a central entity to control and coordinate. It uses a new connection type to connect peers together. The new connection allows the requests (queries) to be passed in only one direction. Our scheme manages those types of connections so that, eventually, contributors become more close to each other in the network, and free riders become isolated. There exist some other studies which also focus on modifying P2P topology such as [16 18,29,30]. However, these works aim to solve the topology mismatching problem and improve the search quality; they do not attack 4

21 the free riding problem directly. In [16], Liu et al. proposed a solution called the Adaptive Overlay Topology Optimization (AOTO) to optimize inefficient overlay topologies for improving P2P search and routing efficiency. In another work [17], Crameret et al. also aimed to create a topology refinement by modifying the bootstrapping mechanism in the P2P network. In [18], Singh and Haahr proposed to modify the P2P network topology so that peers with similar properties become close to each other. Similarly, in [29], Cai and Wang proposed a two-layer (neighbors and friends) unstructured P2P system for better keyword searches. The neighbors overlay is created according to network proximity while the friends overlay is built according to the online query activities. In order to increase the search quality, they try to avoid the free riders in the system while routing the queries. Primarily, the friend overlay is used to route the queries. Because, the friends overlay is constructed in such a way that free riders can not be friends of any peer. However, in their system any peer, including free riders, may issue queries to the system which allows free riders to use the network resources. Chawathe et al. focused on scalability problem in unstructured P2P networks and applied dynamic topology adaptation [30]. They specifically aimed to match the query capacity of the peers with the routed queries to avoid the peers become overloaded by high query rates. 3 P2P Connection Management Protocol In this section, we first describe our motivation and highlight the benefits of our approach through a simple analytic evaluation. We then give the details of our two new connection types and the connection management protocol, that are proposed to control the connections between contributors and free riders. 3.1 Our Approach and Motivation P2P network topology affects the propagation of queries, the quality and quantity of search results, and the overhead imposed on the underlying physical network. Therefore, the connections among peers should be carefully controlled and managed. However, in current unstructured P2P networks, peers can try to connect to any other peer, and they can refuse any connection request to them. Each peer has equal right to do so, independent of their contribution level. Moreover, each peer can use all of its connections to send its queries. In our work, we change these two properties of unstructured P2P network protocols to create an incentive for cooperation and to discourage free riding. 5

22 First, instead of a single connection type that exists in P2P networks to send and receive queries, we define two connection types: IN and OUT connections. IN connections are used to receive queries and to reply them (i.e., provide service). OUT connections, on the other hand, are used just to send queries and to receive replies (i.e., request service). By using two types of connections, we can now differentiate and control service request and service provision separately. Second, we propose a P2P Connection Management Protocol (PCMP) to establish and release these two types of connections. The protocol considers the peer contributions while establishing and releasing connections. Hence free riders can be disconnected from contributing peers and even get isolated sometimes. In this way, the associated problems with free riding can be alleviated. Moreover, contributing peers may establish connections to not free riders, but to other contributors and therefore the number of contributors in their search horizon can be increased. Thus, contributors can have better chance to get query hits and downloads. We foreseen several benefits of applying our protocol. The connectivity of free riders to the contributing peers can be reduced; in some situations, free riders can be totally isolated from the contributors. Furthermore, the connectivity among contributor peers can be increased. Also, the workload of a contributor peer can be reduced, since it will not serve many free riders anymore. As a result, better scalability and robustness can be achieved in the P2P network, since the querying overhead on contributor peers due to free riding can be reduced. With those benefits, we can see improvement in terms of the following quantifiable metrics: Downloads for contributing peers can be increased; Downloads for free riders can be decreased; Amount of query traffic in the network can be reduced. We now provide a motivational example about how we can improve the performance in terms of some of these metrics in a P2P network using our protocol. The probability of getting a query hit depends on many factors including the popularity of the requested file, the number of files shared by peers, and the number of contributing peers in the search horizon. If we assume even popularity and even number of shared files by each peer, then the number of contributing peers in the search horizon will be the factor determining the hit probability of a query. Therefore, increasing the number of contributors in the search horizon is important for receiving better service from the P2P network. In order to calculate the number of contributors that a contributing peer s 6

23 Fig. 1. The relationship between contributors (Cont.) and free riders (FR) at different levels. query can reach, we first do following assumptions. In a P2P network there are contributors and free riders. A peer is considered as a free rider if it does not share any files at all. On the other hand, a peer is a contributor if it shares any number of files. A Gnutella-like protocol is used for the query dissemination with the time-to-live (TTL) value set to m. Each peer in the network has n one-hop neighbors on the average. The number of peers in the network is so large that the path followed by a flooded query constitutes a tree, not a graph. In other words, a query reaches distinct peers at each hop while getting flooded from one hop to the next. A contributor has p number of contributor neighbors and n p number of free rider neighbors. Similarly, a free rider peer has q number of contributor neighbors and n q number of free rider neighbors. Let X i denote the number of peers that are i hops away from the querying peer. We also say X i is the number of peers at level i. X i can be computed easily. X i = n(n 1) i 1, i 1 (1) Some of these X i peers are contributors and some are free riders. Let C i be the number of contributors and F i be the number of free riders at level i. Thus, X i = C i + F i. As we deal with a contributor as the originator of the query, C 0 = 1, C 1 = p, and F 1 = n p. We will compute C i in a recursive manner. Figure 1 shows the relationship between contributors at level i 2, i 1, and i. If we assume that C i 2 is known then F i 2 can be calculated as F i 2 = X i 2 C i 2. 7

24 Upon receiving the query, C i 2 number of contributing peers at level i 2 will forward it to their contributing neighbors (whose count is denoted with C1 i 1 ) and to their free riding neighbors (whose count is denoted with F1 i 1 ) at level i 1. Similarly, F i 2 number of free riding peers at level i 2 will forward the query to their contributing neighbors (C2 i 1 ) and to their free riding neighbors (F2 i 1 ) at level i 1. As indicated in Figure 1, we can compute the number of contributors at level i using the number of contributors and free riders at previous levels i 1 and i 2. Each of the C1 i 1 contributing peers at level i 1 will forward their query to p 1 contributors 2. Then we obtain the following recursive relationship for the number of contributors at level i: C i = C1 i 1 (p 1) + F1 i 1 (q 1) + C2 i 1 (p) + F2 i 1 (q), C i = C1 i 1 p C1 i 1 + F1 i 1 q F1 i 1 + C2 i 1 p + F2 i 1 q, C i = p(c1 i 1 + C2 i 1 ) + q(f1 i 1 + F2 i 1 ) (C1 i 1 + F1 i 1 ). We have the following equations: C1 i 1 + C2 i 1 = C i 1, and F1 i 1 + F2 i 1 = X i 1 C i 1 ; and C1 i 1 + F1 i 1 = C i 2 Y i 2. Here, Y i is the number of neighbors that will receive a query originated or forwarded by a peer i. If the peer is the query originator, i.e. i = 0, the number of neighbors to whom the query will be forwarded is n. Otherwise, if the peer is a query forwarder, the number of neighbors to whom the query will be forwarded is n 1. In short, if i is 0 then Y i is n, otherwise Y i is n 1. Now, the equation that gives the number of contributors at level i becomes: C i = pc i 1 + q(x i 1 C i 1 ) Y i 2 C i 2, i 2 (2) As mentioned before, if the originator of the query is a contributor, C 0 = 1 and C 1 = p. As a result, the total number of contributors that will receive the query issued by a contributor is: m m C = C i = p + (pc i 1 + q(x i 1 C i 1 ) Y i 2 C i 2 ), m 2 (3) i=1 i=2 2 We have p 1 not p because, those forwarding peers have a contributor parent that is also a neighbor of them. 8

25 We can use this recursive formula to compute the number of contributors for various settings of the parameters m, n, p, and q. For example, in a P2P network, each peer, a contributor or a free rider, has 2 contributing neighbors and 3 free riding neighbors. That is, n = 5, p = 2, q = 2, and m = 5. Using Equation 3, the number of contributors that a contributing peer s query can reach is computed as 692. If we can control and modify the connections in this network (what we aim with our approach) so that each contributor has 4 out of its 5 neighbors as contributors (p = 4), then the number of contributors that will receive the query message issued by a contributor would be If we can totally isolate free riders, no free rider will have a connection to a contributor and vice versa. This means, p becomes 5, and q becomes 0. In this case, the number of the contributors that will receive the query would be These examples show that we can improve the number of contributors in a search horizon of a contributing peer so that the peer can get better search quality. This is the main motivation for our approach. After searching the network and receiving the query hits, a peer requests download from one of the source peers. However, source peers are subject to high number of download requests and since the upload capacity is limited, they can refuse some of the download requests. Therefore, receiving a query hit does guarantee a successful download. Assume that on average a contributor can upload U number of files simultaneously at maximum, and the number of simultaneous download requests that arrive to this contributor is D. Sometimes, contributors can have much more download requests (D) than their upload capacity (U). In that case, when D is larger than U, a contributor will refuse a download request with a probability P(refuse) = 1 U/D. As the ratio of free riders in a P2P network becomes greater than that of contributors, then most of these requests will belong to the free riders. As stated above, we aim to reduce the arrival of download requests from free riders. Therefore, we expect a reduction in P(refuse) for the requests coming from contributors. Hence, we expect an increase in the downloads that contributors can achieve. An important issue in realizing our approach is to identify free riders efficiently and correctly. For this, we use a heuristic approach which depends on mutual exchanges of files and query hits between a pair of peers. Based on these exchanges, peers try to identify free riders and contributors. After then they take necessary actions to modify their connections. 9

26 Fig. 2. A general P2P connection between two peers, which enables both of them exchange all types of P2P messages. 3.2 A New Connection Type: One-Way Request Connections In the current unstructured P2P networks like Gnutella, a connection established between a pair of peers is used to exchange all types of P2P protocol messages in both directions including Queries, Query Hits, Pings and Pongs (Figure 2). PCMP modifies this assumption by proposing a new P2P connection type called One-Way-Request Connection (OWRC). As seen in Figure 3, an OWRC between two peers is still a TCP connection and can carry messages in both directions. However, there is a restriction on what types of messages can be carried in which direction of the connection. The connection is called one way because it can transfer requests in only one direction. In other words, over any OWRC the requests (Query, Ping) can only travel in one direction and the replies (Query Hit, Pong) can only travel in the other direction. Such a connection cannot be used to send and receive all kinds of protocol messages in both directions at the same time. The restrictions on the type of messages and their directions are enforced at the application level by PCMP. In Figure 3, one end of the OWRC can be considered a requester (Peer A) and the other end as a responder (Peer B). The requester sends Query and Ping messages and receives the corresponding Pong and Query Hit messages via the OWRC. A responder, on the other hand, receives Query and Ping messages and replies with Query Hit and Pong messages through the same OWRC. In the rest of the paper, we will call such an OWRC an OUT-connection at the requester end and an IN-connection at the responder end. Hence, in Figure 3, peer A has an OUT-connection and peer B has an IN-connection. We will also say that peer A has an OUT-connected peer, which is peer B. And peer B has an IN-connected peer, which is peer A. If we would like to transfer requests from the other direction as well, from B to A, we need to establish another OWRC directed from B to A as depicted in Figure 4. However, we stress again that these connections are logical and can be implemented on top of either one or two TCP connections. A P2P network established using OWRCs can be modelled as a directed graph. A directed arc represents an OWRC: the tail of the arc has the peer that considers the connection as an OUT-connection, and the head of the arc (i.e. the pointing part) has the peer that considers the connection as an IN-connection. Hence the requests can flow along the direction of the arcs. 10

27 Fig. 3. An OWRC between two peers, which limits the direction and the types of P2P messages exchangeable. Fig. 4. Two OWRCs between two peers, which enable each peer to request service from the other. Fig. 5. A directed graph representation of a network consisting of OWRCs. Figure 5 shows an example model of a P2P network consisting of OWRCs. Here, peer A has 6 neighbors. It has four OUT-connected neighbors (B, D, F, G) and three IN-connected neighbors (C, E, G). In other words, the INconnections of A are {C, E, G}, and the OUT-connections of A are {B, D, F, G}. When Peer A would like to search the network it can submit the Query only to its OUT-connected neighbors, namely B, D, F, and G. It will process the Queries only coming from its IN-connected neighbors (C, E, G). If it receives any Query from OUT-connected neighbors it drops the request. The details of a peer interaction with the PCMP are explained in Section 3.5. We believe that peers would like to minimize the number of IN-connections, and they would like to maximize the number of OUT-connections. Because, IN-connections require a peer to process incoming Query and Ping messages, forwarding them and returning any replies to the originator. In contrast, more OUT-connections will help a peer to reach more other peers and increase the 11