Papers Reviews

Property	Change Add	Value
Conference and track		10th IEEE Global Internet Symposium 2007 - 10th IEEE Global Internet Symposium 2007
Authors		Name ID Affiliation Email Country Email Move authors Delete Ching-Ju Lin 149423 National Taiwan University cjlin@cmlab.csie.ntu.edu.tw Taiwan Yi-Ting Chang 194779 National Taiwan University seashell@cmlab.csie.ntu.edu.tw Taiwan Shuo-Chan Tsai 194783 National Taiwan University r93922069@ntu.edu.tw Taiwan Cheng-Fu Chou 168811 NTU ccf@csie.ntu.edu.tw Taiwan
Presenter		Ching-Ju Lin
Registration code
Title		Distributed Social-based Overlay Adaptation for Unstructured P2P Networks
Abstract		The widespread use of Peer-to-Peer (P2P) systems has made multimedia content sharing more efficient. Users in a P2P network can query and download objects based on their preference for specific types of multimedia content. However, most P2P systems only construct the overlay architecture according to physical network constraints and do not take user preferences into account. In this paper, we investigate a social-based overlay that can cluster peers that have similar preferences. To construct a semantic social-based overlay, we model a quantifiable measure of similarity between peers so that those with a higher degree of similarity can be connected by shorter paths. Hence, peers can locate objects of interest from their overlay neighbors, i.e., peers who have common interests. In addition, we propose an overlay adaptation algorithm that allows the overlay to adapt to P2P churn and preference changes in a distributed manner. We use simulations and a real database called Audioscrobbler, which tracks users' listening habits, to evaluate the proposed social-based overlay. The results show that social-based overlay adaptation enables users to locate content of interest with a higher success ratio and with less message overhead.
Topics		P2P networking and overlay networks; Content networking (caching, content distribution, content routing, content services, load balancing, etc.)
Session		(not assigned to a session)
TPC group		none
Status		accepted Authors were notified March 23, 2007.

5 Reviews

Assign another review

Review 1

Reviewer: Azer Bestavros; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 9, 2007

Strength (Explain the strength of this paper (2-4 lines))

The authors argue that a cluster-based construction of content distribution overlays, based on user preference similarity, boosts their performance. This statement is important as most of the references in the literature focus on enhancement to overlay network performance based on physical network constraints, ignoring the similarity of user request profile (and consequently the proximity to the requested objects).

Strengths/Contributions:

(1) The authors study the performance of Semantic-based overlay that are constructed by a central authority or in a distributed fashion (using random walks) in comparison with those constructed by local heuristics. The experimental space is large; results reported include: (a) Centralized Semantic-based networks outperform (in terms of success and precision rate) decentralized Semantic-based networks, which in turn outperform overlay networks constructed using local heuristics. (b) This performance comes with a high cost of message exchange for the centralized approach and a significantly lower cost for the decentralized versions.

(2) Under their setting free riding is reduced (by not contributing content the user similarity is miscalculated).

(3) The proposed framework is also shown to be able to adapt to churn.

Overall, the paper is well written

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

There is not enough validation why the described quantifiable measures of performance are the ones to consider.

Details (Discuss the merits and provide suggestions for improvement)

The authors should be aware of and should refer to "Implications of Selfish Neighbor Selection in Overlay Networks" Nikolaos Laoutaris, Georgios Smaragdakis, Azer Bestavros and John W. Byers. IEEE INFOCOM 2007. In this work, user preference is also taken into account in the neighbor selection process, leading to highly optimized overlay networks. They are also encouraged to check with the authors of the above work for newer results and implementations.

It would have been nice to see some modeling and analysis in the paper. I realize that it is not straightforward, but in the spirit of suggesting ways teh work could be improved, I think this is a good direction (also as the above reference and follow-up work has done, it seems that some analysis is possible).

Review 2

Reviewer: Anup Basu; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 9, 2007

Strength (Explain the strength of this paper (2-4 lines))

Considers clustering of peers based on preferences.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The similarity function seems too simple.

Details (Discuss the merits and provide suggestions for improvement)

Probably more sophisticated multidimensional clustering could be used.

Review 3

Reviewer: Sonia Fahmy; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 9, 2007

Strength (Explain the strength of this paper (2-4 lines))

The paper discusses the interesting problem of developing a distributed algorithm for forming unstructured P2P networks based on the similarity of user interests. A measure of similarity is used in conjunction with node IDs to form and adapt the overlay topology. Simple simulation results are given.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

It would be interesting to combine this approach with a network-aware (e.g., latency aware method) rather than a random method. The simulation scenario is not very representative in some aspects (see below).

Details (Discuss the merits and provide suggestions for improvement)

Overall, I enjoyed reading the paper and found the general idea interesting.

The paper mentions that structured P2P networks have a number of limitations without citing/giving supporting evidence -- an NSDI 2005 paper gives interesting results on this, and this paper can benefit from that information.

It is not clear how BRITE was used to generate the topologies and how hosts were attached -- are these AS-level topologies only? It would be more accurate to use router-level ones with delays assigned on links.

A discussion on how the results in section IV would scale if the number of fans (users) increased beyond 1355 would be interesting.

The churn scenario is not representative of churn in real networks. It would be better to use a trace for that.

Some minor problems:

Page 3: we merges -> we merge Page 3: which connects two peers named the consecutive identifiers -> this statement needs to be rewritten Page 3: based the objects -> based on the objects {age 3: It then builds temporary overlay links with those peers to connect ..., exchange information (should be "exchanges").... and compile its buddy (should be "and compiles"). Page 4: "the weak graph" should be precisely defined

Review 4

Reviewer: Christos Gkantsidis; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 10, 2007

Strength (Explain the strength of this paper (2-4 lines))

The basic idea of connecting users with similar interests is interesting, but definitely not new. However, I really liked that the authors have used real data to evaluate their ideas.

The evaluation is pretty thorough for a short paper. (However, the number of users, interests, etc. is small.)

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

It is not clear how the basic idea will scale to systems with millions of users and 100K of topics of interest. Even computing the similarity between peers may involve exchanging huge vectors in that case.

Admittedly the sample set used to evaluate the ideas is small.

Details (Discuss the merits and provide suggestions for improvement)

THIS WAS SUPPOSED TO BE AN ANONYMOUS SUBMISSION. Please read the instructions carefully before submitting.

The most interesting parts of this paper is the evaluation using real data, and the use of random walks and neighbors to sample random peers.

The observation that random walks and local neighbors give performance close to the optimum (SFN) is interesting. However, it may be an artefact of the small size of the network. You should comment on whether you expect that your observations will be carry over to large networks.

You use the word "distributedly" quite often; however, it is not very common and I would propose that you change your wording.

You need to define "weak graph" (end of 1st paragraph in Sec.IV).

Review 5

Reviewer: Jun Li; completed; last reminded Mar 11, 2007; assigned Feb 15, 2007; completed Mar 13, 2007

Strength (Explain the strength of this paper (2-4 lines))

This is a solid paper with a novel idea.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

No significant technical weakness was found.

The paper does not follow the submission format requirements to stay anonymous and use 11-pt font.

Details (Discuss the merits and provide suggestions for improvement)

This paper has a very good motivation in clustering peers with similar social interests. It presents clearly what is the state of the art, and what's the knowledge gap. The solution proposed is clear and sound, and the paper is very easy to follow. It also provides incentives for peers to participate in such an overlay.

I only have some minor points regarding this paper, namely:

(1) Is it always true that peers are similar should be next each other? How about nodes with totally different but complementary contents? I can imagine scenarios that a peer with mostly music files may want to be close to a peer with mostly video files and begin to download videos.

(2) The taste of a peer can change all of sudden, but because the peer has to gradually change the makeup of its files, the real interest of a peer may not be accurately captured based on its local files.

(3) How accurate is it to use the files of a peer as an indicator of the peer's profile? Why not let the user to express what files he wants and what files he provides, and let every peer to be close to those peers that provide what the peer wants?

Property	Change Add	Value
Conference and track		10th IEEE Global Internet Symposium 2007 - 10th IEEE Global Internet Symposium 2007
Authors		Name ID Affiliation Email Country Email Move authors Delete Börje Ohlman 123098 Ericsson borje.ohlman@ericsson.com Sweden Anders Eriksson 148493 Ericsson Research anders.e.eriksson@ericsson.com Sweden
Presenter		?
Registration code
Title		Dynamic Internetworking Based on Late Locator Construction
Abstract		The legacy Internet technology is optimized for a semi-static inter-domain topology. Mobility or multihoming is often handled by extending the legacy technology with new protocols. This paper describes a novel internetworking architecture with native support for a highly dynamic edge topology. The architecture separates between a rather static core network on the one hand, and on the other hand edge networks forming an edge topology that may change on a short timescale due to mobility or re-homing events. To avoid indirections via mobility agents, the source host addresses the destination host directly with a hierarchically structured locator. The name to locator resolution is based on a novel mechanism that constructs the host locator on-demand to describe the current internetwork path from the core to the host (late locator construction). This enables the resolution of the host name into a hierarchical and topologically significant host locator also in a highly dynamic edge topology where the path to the host traverses several moving and multihomed networks. Mobility and multihoming are treated as closely related concepts, and the same name resolution mechanism is used for both.
Topics		Handling Internet dynamics/heterogeneity (by applications and/or the network); Routing (unicast, multicast, anycast, etc.); The Internet and mobility/mobile devices
Session		(not assigned to a session)
TPC group		none
Status		accepted Authors were notified March 23, 2007.

5 Reviews

Assign another review

Review 1

Reviewer: Taieb Znati; confirmed; last reminded Mar 20, 2007; assigned Feb 15, 2007

Review 2

Reviewer: Henning Schulzrinne; completed; assigned Feb 15, 2007; completed Mar 5, 2007

Strength (Explain the strength of this paper (2-4 lines))

The paper proposes a new edge network architecture that uses source routing to avoid mobility anchor points. The approach could potentially address network mobility.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The paper only presents an architecture and unfortunately does not contain any performance or scalability analysis, beyond a bit of handwaving along the way.

Details (Discuss the merits and provide suggestions for improvement)

The lack of evaluation makes it difficult to evaluate and compare it to other efforts, such as the micro-mobility protocols (such as HAWAII and cellular IP, to name two of many) that were developed a few years ago and which are not mentioned?

Also, since edge networks have to run a routing protocol, this doesn't appear to differ all that much from existing mechanisms, with the only major advantage being the construction of the end identifier.

It would be helpful to provide hints as to the anticipated mobility rates of edge networks. Would the system likely to work if edge network connections changed every few seconds, such as might be the case for ad-hoc network consisting of mobile vehicular networks, or only if the topology is stable over longer time periods?

As with many architecture papers, it is difficult to tell how the architecture was arrived at. For example, why weren't simpler two-layer architectures consisting of a core attachment identifier and a local network identifier (e.g., similar to GSE/8+8) not considered?

The paper seems to want to address two concerns, namely avoiding mobility agents and routing updates. However, these are largely orthogonal, as there are several other, simpler mechanisms that avoid mobility agents, such as session-layer mobility or FMIP. (Like other proposals that avoid anchor points, the paper does not address the issue of simultaneous mobility if both MH and CH are mobile.)

The principal advantage appears to be multi-homing. Again, there have been numerous multi-homing proposals that would appear to be applicable to the problem at hand, but the paper makes no attempt to differentiate itself from earlier such efforts.

Given the large number of mobility architectures proposed, the paper falls short in motivating its rather complicated design, requiring significant network layer changes. It would make comparison easier if the proposal was separated into components that can be compared and evaluated individually. For example, the impact on backbone routing does not appear to depend on the details of source route construction.

Review 3

Reviewer: Daniel Zappala; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 8, 2007

Strength (Explain the strength of this paper (2-4 lines))

The main strength of this paper is that it explores a new network architecture that integrates support for mobility at the network edges. This idea has been around for a while, but has recently gained some traction, with increasing interest in exploring alternative architectures. The architecture tries to limit the impact of mobility on the core of the network, while avoiding indirection.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The authors do not make a convincing case for why this architecture is needed, making only a passing reference to mobility and multihoming, without discussing these problems in detail.

The discussion of the architecture is hard to follow -- the authors do not clearly explain the "big picture" of how it works and do not adequately walk through examples.

Details (Discuss the merits and provide suggestions for improvement)

I'm intrigued by the potential of this paper. The authors claim to have a novel architecture that seamlessly supports mobility, with little impact on the core of the network. The architecture provides direct routes to mobile nodes, rather than requiring indirection through a home agent. The result of this design decision is a dependence on "attachment registers", which track the current location of mobile nodes. Each node must detect when it moves, discover its new route to the core, and then keep its register up-to-date with its current location.

Unfortunately, there are several problems with the paper that could hinder its impact at a conference.

First, the authors do not discuss any motivating examples that would explain their design choices. For example, are the edge nodes a mobile ad hoc network? Explain how the network would be used, and this may drive the need for a new architecture and a particular design.

Second, the overall architecture is not clearly explained. The authors describe various details in the paper, but do not tie it together well enough so that the reader can understand how it works. Several times the authors refer to the steps in Figure 1, without explicitly detailing what occurs at each of these steps. It is not clear exactly what the "attachment registers" store -- a full locator to the host or just the identifier for the next hop toward the host/network?

Third, there is no evaluation of even a part of the architecture. There is some discussion of scalability concerns, but it would really help if the authors picked one aspect of the architecture and made a first attempt at determining how well it would work.

Other concerns:

Related work is not adequately covered -- a quick read of citation [5] shows a paper from last year's conference that covered this area in much more detail.

Regarding the design itself, the architecture appears to depend heavily on iterative lookups at the "attachment registers". The paper could benefit from a deeper discussion of how these registers work and how quickly they can be updated, given frequently moving hosts. This system seems to be similar to a very dynamic DNS -- how does it differ, and how could your particular design better handle the dynamics of a highly mobile edge?

Review 4

Reviewer: Peter Reiher; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 9, 2007

Strength (Explain the strength of this paper (2-4 lines))

This paper proposes a different method of performing Internet routing that might be better suited to mobile machines and multihoming. It uses standard Internet routing at the core, but introduces new, potentially more flexible, mechanisms at the edges, where multihoming and mobility typically occur. The ideas are interesting and the presentation clear.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The authors have perhaps not sufficiently examined some of the hard issues related to the basic ideas, making it difficult to determine if the approach is workable or infeasible.

Details (Discuss the merits and provide suggestions for improvement)

I found this an interesting paper to read and one that provoked a certain amount of thought. I think it fits well into the goals of the symposium and is likely to be of interest to most of the attendees.

There were a number of key issues that either weren't addressed or did not receive the attention they needed. First, nothing is said about routing security. The problems existing routing techniques have with security suggest that we should not seriously consider any other technique without understanding its security problems and implications. Also, there was not sufficient discussion of issues of reliability in the data structures used to store routing information. Just saying "use DHTs" is not enough.

At first, it seemed the authors were not really going to address the effects of routing changes on ongoing sessions. They did get around to doing so, but it would be good to make it clearer earlier that they were going to handle them reasonably. Also, it's not clear that the approach they suggest, querying the LCS for a new GL, can be done fast enough to keep sessions alive. I'd like to see the authors address that issue.

Ultimately, the core of the approach amounts to partial source routing. There are known problems with this idea, particularly with the per-packet overhead of carrying the route along with you. The authors need to address this issue better.

The discussion on reducing the size of the host GL is too brief and lacks sufficient detail. It's not clear to me that the IPv6 prefix delegation approach easily maps to their problem. This is an important issue that the authors need to address.

The first paragraph of the "Robustness and disconnected operation" section does not make clear how the LCSR detects inconsistencies and purges stale registrations.

The authors claim their scheme has the advantage that the core routing system does not receive (and thus need not keep track of or do anything else about) routes from the edge. However, it does need to host all the ARs, and they must be updated. It's not clear to me that this is an improvement, and, depending on details, it might even be worse.

Finally, a style quibble. "Usage" is just a more pretentious form of "use." I'd recommend always using "use," not "usage."

Review 5

Reviewer: James Sterbenz; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 11, 2007

Strength (Explain the strength of this paper (2-4 lines))

A reasonable architectural proposal for identifying and binding to mobile hosts in the Internet. This paper is likely to generate good discussion at GI, and will be of interest since it is part of the large EU Ambient project.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

This is only an architecture paper with no analysis, although the authors do consider and qualitatively discuss a number of important issues. Given that this is part of an ongoing EU project, the status of this work with respect to a prototype should have been given.

Details (Discuss the merits and provide suggestions for improvement)

The paper is well written and organised; a few minor typos are noted below.

Presumably, you are assuming a relatively low number of edge hops to the core to keep the LCSR efficient? It would be nice if you would explicitly state your assumptions in this area and note their impact.

You might want to note some similarity in the way reverse DNS lookup is done.

How do you propose to handle Policy and QoS routing? Is is merely a matter of building conventional mechanisms into a link-state LCSR, or is there more to is?

Since this is part of the Ambient project, I assume you are building a prototype? You should describe the status of this work. If it *has* already been implemented, then you should briefly describe your experience.

Replace "(EID" with "(EID)" Replace "to compresses the" with "to compress the"

The references are weak, but I realise you have limited space.

Property	Change Add	Value
Conference and track		10th IEEE Global Internet Symposium 2007 - 10th IEEE Global Internet Symposium 2007
Authors		Name ID Affiliation Email Country Email Move authors Delete Ingmar Baumgart 146352 Universität Karlsruhe (TH) baumgart@tm.uka.de Germany Bernhard Heep 204430 Universitaet Karlsruhe heep@tm.uka.de Germany Stephan M Krause 204432 Universitaet Karlsruhe stkrause@tm.uka.de Germany
Presenter		Stephan Krause
Registration code
Title		OverSim: A Flexible Overlay Network Simulation Framework
Abstract		A fundamental problem in studying peer-to-peer networks is the evaluation of new protocols. This paper presents OverSim, a flexible overlay network simulation framework based on OMNeT++. It was designed to fulfill a number of requirements that have been partially neglected by existing simulation frameworks. OverSim includes several structured and unstructured peer-to-peer protocols like Chord, Kademlia and Gia. These protocol implementations can be used for both simulation as well as real world networks. To facilitate the implementation of additional protocols and to make them more comparable OverSim provides several common functions like a generic lookup mechanism for structured peer-to-peer networks and an RPC interface. Several exchangeable underlay network models allow to simulate complex heterogeneous underlay networks as well as simplified networks for large-scale simulations. We show that with OverSim simulations of overlay networks with up to 100,000 nodes are feasible.
Topics		P2P networking and overlay networks; Traffic measurement, analysis, modeling, and visualization
Session		(not assigned to a session)
TPC group		none
Status		accepted Authors were notified March 23, 2007.

5 Reviews

Assign another review

Review 1

Reviewer: Craig Partridge; completed; assigned Feb 15, 2007; completed Mar 5, 2007

Strength (Explain the strength of this paper (2-4 lines))

Presents a simulator for overlay networks that includes the ability to include emulated hosts (and thus run real code in the simualted environment).

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

Every so often the explanation falters -- some features are explained without motivating why they exist -- some tests are done without explaining why they are the right tests.

Details (Discuss the merits and provide suggestions for improvement)

Mostly this is a fine paper. My concerns are mostly a matter of presentation.

In section II, it would work better, I think, to crisply explain what features an ideal simulator has and then discuss how each work fits that ideal. For instance, some simulators apparently lack good logging/statistical output [apparently an ideal goal], others have scalability issues, etc. Rather than force the reader to puzzle these out from what the paper chooses to highlight, let's be explicit.

The paper uses three different notations for numbers: a European notation (10.000) a US notation (10,000), and what I think is IEEE notation (1000). Using one notation throughout would help. (Figure 3, for instance, uses one notation on the X-axis and another on the Y-axis).

Section IV, C. I agree that sharing certain services to prevent replication matters, but I didn't ever get a good understanding of why these particular services were the right one's to share -- indeed, some of the features -- such as providing random addresses in the network, seemed central to how different overlay networks might wish to build themselves and thus not be something that should be shared?

Section V, telling me that a network has 20 backbone and 20 access routers tells me very little about the complexity of the topology -- was this, perhaps, a simple dumbell or a ring, or a random laydown?

Section VI -- I like the idea of validating but the text didn't explain why this validation was the right one -- does Chord have a known failure mode of collapsing into multiple rings in simulators???

Review 2

Reviewer: Michael Devetsikiotis; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 8, 2007

Strength (Explain the strength of this paper (2-4 lines))

Well motivated, written and illustrated paper in timely area, highly relevant to the GI workshop.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The network model and the "DHT test application" are overly simplistic. The "validation" portion is somewhat weak.

Details (Discuss the merits and provide suggestions for improvement)

Well written manuscript, structured and illustrated in a way that makes it useful. The subject area is very timely and relevant to the GI workshop.

Overview and references are quite thorough. The progression from overview to "requirements" to description of the new simulator, and ultimately to a quantitative comparison, is logically appealing and increases the practical usefulness of the paper. The details about the simulation are very welcome as they make the results reproducible,

The network model and the "DHT test application" are simplistic but both can be expanded (especially interesting would be how to model the application layer).

Useful discussion in a timely area, makes it a good paper for the workshop.

Review 3

Reviewer: George Stamoulis; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 9, 2007

Strength (Explain the strength of this paper (2-4 lines))

The paper describes the implementation of an innovative simulator (OverSim) that combines many features of existing simulators. This simulator is experimentally compared with P2PSim in terms of scalability and correctness (in terms of path length per lookup of a random node) and proved to be scalable and correct. The paper is well-written and structured. It is a description of solid development work, rather than a typical research paper.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

As mentioned in Sections VI and VII, the simulator has not been completely verified yet, neither it has not been compared with other simulators than P2PSim. The level of conceptual innovation of the simulation should be further clarified. By the way it is now presented, it seems that OverSim is mostly a combination of known approaches for the various aspects.

Details (Discuss the merits and provide suggestions for improvement)

This paper describes a P2P simulator that supports multiple overlay protocols for structured Chord, Kademlia, Koorde, Broose and unstructured peer-to-peer protocols like GIA, as well as multiple underlay network protocols. These underlay network models can effectively describe a network as simple as a delay graph, or as complex as a per device OSI protocols specification. The protocol combines many important features from existing P2P simulators and it emphasized in the underlay network modelling as well.

Experiments may be both real-time and non real-time. An important feature is that it is possible that experimental P2P nodes interact with real P2P nodes and protocols in the case of real-time. The scalability of the approach as compared with P2Psim is quite good in terms of the number of nodes that can be simulated. The simulator scales linearly with the number of nodes w.r.t. memory usage.

Specific comments -----------------------

Mention explicitly that experiments using OverSim can run in non real time, which is only implied currently in the paper.

Network emulation is an important feature of this simulator. It should be described in more detail. In general, the authors should emphasize more on the description of the innovative features and parts of OverSim.

The applicability of Simple model (Section IV.B) should be further explained. To which case of networks does it correspond ?

In Section V, in the IPv4 simulation run, what are the other characteristics of the underlay network (link rates, delay etc.)?

Review 4

Reviewer: Jorg Liebeherr; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 11, 2007

Strength (Explain the strength of this paper (2-4 lines))

The paper presents describes an implementation of a simulation tool for overlay protocols. The advantage of the tool is that it can support a large number of nodes. I believe that a demonstration of the simulator at the conference will find a lot of interest.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

While the developed tool is highly useful to do networking research, the tool itself is not necessarily a contribution to networking research. The paper may be more suitable for a conference that is devoted to performance evaluation methods and tools.

Details (Discuss the merits and provide suggestions for improvement)

Evaluation tools, such as network measurement systems and simulators, are essential for making new discoveries and contributions in networking research. Once can make a case that a simulation tool itself is a contribution based on the following:

1. The tool enables new discoveries in networking research, that cannot be made with existing tools. 2. The tool itself as a software artefact presents a contribution in the development of software simulators.

Regarding the first point, the paper presents evidence that it can support many more nodes than an existing simulators. Yet, this property is not exploited to make new findings or provide new insights about overlay protocols. Regarding the second point, from a software engineering or simulator design perspective, the OverSim tool is a case study of the OMNet++ simulation environment, but not a novel approach to the problem of simulating large networks.

I believe that a demonstration of the simulator will find a lot of interest at the conference. On the other hand, the paper does not make a strong case that it advances the state-of-the-art of network simulators.

Other comments:

- Some of the points in Section III (Requirements) could be clarified. What is the rationale for setting "10,000" as the lower bound for a "large simulation". Why is "1,000" too low or "1,000,0000" too large?

- It is not clear to me that the limitations of P2Psim are principal, or the result of an inefficient implementation (e.g., supoptimal management of memory).

Review 5

Reviewer: Ljiljana Trajkovic; completed; last reminded Mar 11, 2007; assigned Feb 15, 2007; completed Mar 12, 2007

Strength (Explain the strength of this paper (2-4 lines))

The authors describe a new network simulation tool for P2P networks.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The ideas presented in the paper are at the preliminary stage. The authors need to provide a more thorough validation and performance evaluation of the proposed tool.

Details (Discuss the merits and provide suggestions for improvement)

The authors present and interesting and useful simulation tool. Nevertheless, the presented work is at the preliminary stage and the proposed simulation framework needs to be validated using larger P2P network test cases. Its performance should also be compared to the existing P2P simulation tools based on the ns-2 simulator.

The paper needs to be carefully rewritten. The authors are strongly advised to use a more formal style of presentation better suited for technical and research papers.

Additional reference: A. Dufour and Lj. Trajkovic, ``Improving Gnutella network performance using synthetic coordinates,'' Proc. The Third Int. Conf. on Quality of Service in Heterogeneous Wired/Wireless Networks (QShine 2006), Waterloo, ON, Canada, Aug. 2006.

Property	Change Add	Value
Conference and track		10th IEEE Global Internet Symposium 2007 - 10th IEEE Global Internet Symposium 2007
Authors		Name ID Affiliation Email Country Email Move authors Delete Krishna Puttaswamy 157492 University of California, Santa Barbara krishnap@cs.ucsb.edu USA Ben Y. Zhao 118980 University of California at Santa Barbara ravenben@cs.ucsb.edu USA
Presenter		Krishna Puttaswamy
Registration code
Title		A Case for Unstructured Distributed Hash Tables
Abstract		Structured peer-to-peer overlays support compelling applications such as large-scale file systems and distributed backup using the distributed hash table (DHT) interface. While unstructured file-sharing systems continue to flourish, wide adoption of structured applications has been elusive. We explore an alternative path to deployment of these applications by asking the question, can structured applications be run on top of unstructured overlays? We build an unstructured distributed hash table (UDHT) on top of state of the art search and topology management mechanisms, and evaluate whether it can sufficiently emulate properties of DHTs to support structured applications.
Topics		P2P networking and overlay networks; Novel applications and new paradigms (telephony, streaming media, etc.); Handling Internet dynamics/heterogeneity (by applications and/or the network)
Session		(not assigned to a session)
TPC group		none
Status		accepted Authors were notified March 23, 2007.

5 Reviews

Assign another review

Review 1

Reviewer: Anup Basu; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 9, 2007

Strength (Explain the strength of this paper (2-4 lines))

Suggests the use of unstructured distributed hash tables.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

It is not clear how the performance compares with existing hash table methods.

Details (Discuss the merits and provide suggestions for improvement)

See comments above.

Review 2

Reviewer: Azer Bestavros; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 9, 2007

Strength (Explain the strength of this paper (2-4 lines))

The authors study the deployment of structured applications on top of unstructured overlays. They introduce the notion of UDHT (Unstructured Distributed Hash Table) and evaluate if it can guarantee DHT properties over unstructured overlays. The problem is important as a successful implementation of UDHT will leverage the deployment of (controlled) applications over unstructured overlays.

The main contributions are that random walk are more efficient than flooding (expected - reported in the literature) and that adding more replicas would increase the success rate (expected - reported in the literature).

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The problem is important, but the paper does not really deliver what it promised, i.e., to provide a "novel" framework to create structured applications over unstructured overlays. There is no comparison with DHT in the same setting. Conjecture: Under the same setting, structured overlays may be even better than unstructured overlays.

Details (Discuss the merits and provide suggestions for improvement)

Paper could use some proofreading, e.g., on page 2/Column 1/Paragraph 2: shouldn't "2113 retrieves data stored by node 1023" be "1023 retrieves data stored by node 2113" (?)

Review 3

Reviewer: Christos Douligeris; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 11, 2007

Strength (Explain the strength of this paper (2-4 lines))

The fact of the limited deployment of DHT applications have driven the authors to examine an alternative approach, namely an Unstructured Distributed Hash Table. The argument of the expected performance of DHTs when the number of nodes is under a certain threshold (inverse scaling problem) as well as the potential benefits of an underlying unstructured overlay warrants the feasibility study of the UDHT.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

There is a need of further comparisons and more readable figures.

Details (Discuss the merits and provide suggestions for improvement)

The fact of the limited deployment of DHT applications have driven the authors to examine an alternative approach, namely an Unstructured Distributed Hash Table. The argument of the expected performance of DHTs when the number of nodes is under a certain threshold (inverse scaling problem) as well as the potential benefits of an underlying unstructured overlay warrants the feasibility study of the UDHT.

After a brief introduction to the structured overlay networks, the authors explain their choices with respect to search algorithms and topology adaptation methods. In particular, UDHT relies in search methods previously defined in the context of Gia. The employment of k-random walks significantly reduces the overhead of flooding based approaches. The authors introduce an embedded n-window random walk to avoid redundant paths. In contrast to Gia's biased random walk, where each node maintains a per-query state, UDHT embeds in each query the last n nodes visited, thus avoiding routing loops of n hops or less. A node doesn't have to keep per query state, but the query length is increased, hence the network traffic. UDHT relies on previous work to exploit the heterogeneity, in terms of node capacity, and achieve better performance. Moreover, the employment of one-hop index replication further improves the search effectiveness.

UDHT maintains n (replication factor defined as system parameter) replicas for each object, where n is a system wide parameter determined at start-up. The n peers storing the replicas are referred as replica set for key K. Members of each replica set must maintain indices to all members of the replica set. The authors then explain how the GET and PUT operations of the DHT interface are implemented using random walks. In particular, the GET operation is implemented as a 3-random walk search. The PUT method is implemented by fist issuing a GET. If the object exists, the replica set is returned and the new object is stored in each peer in the replica set. If no replica set is found a new replica set is created by issuing n parallel random walks.

Finally a series of experiments is presented,organized into two sets. The former set evaluates Gnutella's and Gia's ability to locate unpopular objects, while the latter evaluates the effectiveness of UDHT. In short, 1. they show the lookup query success for different number of random walks and depth (fig 2a and 3a). The Gia topology outperforms Gnutella. From the evaluation the authors conclude that their UDHT will enable one-hop replication and use one random walks (static) 2. they show how the replication factor can increase the rate of successful lookup (fig 2b) while decreasing the query overhead and bandwidth requirements(fig 3b and 3c). (static) 3. they show the benefits of increased replication factor under gnutella and skype churn models.(fig 2c) 4. they show the relative costs of put and get methods in terms of message hops. 5. they show the decrease in the effectiveness of UDHT, in terms of successful lookups (fig 4b) and average lookups (fig 4c) as the network size increases. The experiments show that UDHT is effective in smaller networks while in larger networks the effectiveness of random walk search gradually degrades.

Since the authors introduce a new random walk method (embedded n random walk) it would be appropriate to provide a comparison between Gias random walk and UDHT (e.g. in fig3c). Moreover the replication factor could be expressed in terms of a percentage of the node population (RF=1% means that one percent of the population have the given object), to make the comparisons in figures 2a and 3a fair. Finally a graph indicating the delay of queries (time taken for a query from start to finish) would be representative for the effectiveness, from a user's perspective, of UDHT.

Review 4

Reviewer: Sonia Fahmy; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 13, 2007

Strength (Explain the strength of this paper (2-4 lines))

The paper investigates the construction of a protocol layer on top of unstructured P2P networks that provides a DHT-like interface, with the goal of facilitating the deployment of DHT applications. The paper is extremely well-written and does ask a question that has not been explored in the literature. Simulation results show the proposed layer works reasonably well on small to medium scales.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The simulation results can be stengthened in that they currently explore overlay-level metrics only (it would be interesting to know the real delays), and use traces that may not be well-suited to this study (e.g., I believe that the Skype trace used only included supernodes). The simulations can also include comparisons. The presentation of the results can also be improved (see below).

Details (Discuss the merits and provide suggestions for improvement)

Overall, the paper was interesting to read, and acknowledges the scalability issue as an area for future work.

As mentioned above, real performance metrics would be interesting to explore, e.g., real network delays as opposed to application-level metrics only. Comparisons can also strengthen the results.

I find the layout of the figures and the order in which they are referenced rather confusing. Instead of grouping the subfigures into figures in what seems to be an unnatural grouping (it does not correspond to order of references), I suggest making each into a separate figure (which can still be placed 3 per row) and ordering them to correspond to reference order.

Page 3, last paragraph of section III: a new nodes connects -> a new node connects

Review 5

Reviewer: Georg Carle; completed; last reminded Mar 7, 2007; assigned Feb 15, 2007; completed Mar 15, 2007

Strength (Explain the strength of this paper (2-4 lines))

The idea of an unstructured DHT puts together some nice methods. It is really interesting to combine unstructured networks for small-scale and at some point turn more structured as the network gets bigger. However, this point is not yet solved by the authors.

Weakness (Explain the areas in which this paper could be improved (2-3 lines))

The motivation is kind of weak as DHTs are already used in current Peer-to-Peer systems where they make sense. The paper lacks a comparison to other systems. The query hit rates and search path lengths do not seem to be very good.

Details (Discuss the merits and provide suggestions for improvement)

The authors propose an unstructured distributed hash table. The name is a bit misleading as it supports the same API as DHTs but does not necessarily offers the same properties (e.g. false negatives, etc.). The motivation of the work is that unstructured networks are more widely used. One may question this assumption as modern Peer-to-Peer system combine different technologies. In particular, one can doubt that making DHTs unstructured improves usage of DHTs.

The work presented in the paper is interesting. The authors realize put/get functions in their unstructured system and introduce replication. The evaluation is based on models derived from real-world measurement studies, including Gnutalla and Skype churn models.

The advantages of random walk searches (at least to some degree) are based on Power-Law network assumptions. The UDHT seems to try to evolve into the direction of such topologies, but in particular in small scenarios one may wonder if the UDHT will create such topologies and thus, still profit from their properties.

Another problem is the following. Consider that many items are stored. A node may end up in a lot of replica sets each fully connected. Thus the node may be connected with more nodes in the network than desirable.

Remark: Review by Heiko Niedermayer and Georg Carle

	Change Add	Value
Conference and track		10th IEEE Global Internet Symposium 2007 - 10th IEEE Global Internet Symposium 2007
Authors		Name ID Affiliation Email Country Email Move authors Delete Roman Chertov 128836 Purdue University rchertov@cs.purdue.edu USA Sonia Fahmy 6339 Purdue University fahmy@cs.purdue.edu USA Ness Shroff 6199 Purdue University shroff@ecn.purdue.edu USA
Presenter		?
Registration code
Title		A Black-box Router Profiler
Abstract		Simulation, emulation, and wide-area testbeds exhibit different strengths and weaknesses with respect to fidelity, scalability, and manageability. Fidelity is a key concern since simulation or emulation inaccuracies can lead to a dramatic and qualitative impact on the results. For example, high--bandwidth denial of service attack floods of the same rates have very different impact on the different platforms, even if the experimental scenario is supposedly identical. This is because many popular simulation and emulation environments fail to account for realistic commercial router behaviors, and incorrect results have been reported based on experiments conducted in these environments. In this paper, we describe the architecture of a black-box router profiling tool which integrates the popular ns-2 simulator with the Click modular router and a modified network driver. We use this profiler to collect measurements on a Cisco router. Through a careful sensitivity analysis, we demonstrate that routers and other forwarding devices cannot be modeled as simple output port queues, even if correct rate limits are observed. We discuss our future work plans for using our data to create high--fidelity network simulation/emulation models that are not computationally prohibitive.
Topics		Traffic measurement, analysis, modeling, and visualization
Session		(not assigned to a session)
TPC group		none
Status		accepted Authors were notified March 23, 2007.

5 Reviews

Assign another review

Review 1