On Unbiased Sampling for Unstructured Peer-to-Peer Networks

Stutzbach, Daniel; Rejaie, Reza; Duffield, Nick; Sen, Subhabrata; Willinger, Walter

doi:10.1109/tnet.2008.2001730

Cited by 200 publications

(158 citation statements)

References 41 publications

Supporting

Mentioning

155

Contrasting

Unclassified

Order By: Relevance

“…The ion-sampler tool allows us to gather many samples at once by running many walks in parallel. Our extensive simulations [13] show that the technique yields unbiased samples under a wide variety of peer behavior, degree distributions, and overlay construction techniques. For this study, we use 1,000 parallel walks.…”

Section: Capturing Unbiased Samplesmentioning

confidence: 99%

“…Second, the heterogeneous nature of the overlay topology can lead to bias toward high-degree peers, as peers with a high degree are more likely to be discovered. In our prior work [13], we developed a light-weight sampling tool, called ion-sampler, for P2P systems that selects peers uniformly at random. The ion-sampler tool performs a random walk and selects the peer at the end of the walk as a sample.…”

Section: Capturing Unbiased Samplesmentioning

confidence: 99%

“…In static graphs, the Metropolis-Hastings method can be used to weight the walk to result in uniform sampling [19][20][21], eliminating the degree bias. Our prior work has shown [13] that it is also effective in eliminating the degree bias in dynamic networks such as P2P systems.…”

Section: Capturing Unbiased Samplesmentioning

confidence: 99%

“…We also leverage an unbiased sampling technique for P2P system that we developed in our earlier work [12,13] to select a representative subset of peers. This sampling technique is implemented in a light-weight sampling tool, called ion-sampler.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Characterizing files in the modern Gnutella network

2007

View full text Add to dashboard Cite

The Internet has witnessed an explosive increase in the popularity of Peer-to-Peer (P2P) filesharing applications during the past few years. As these applications become more popular, it becomes increasingly important to characterize their behavior in order to improve their performance and quantify their impact on the network. In this paper, we present a measurement study on characteristics of available files in the modern Gnutella system. We develop two new methodologies to capture accurate snapshots of available files in a large-scale P2P system. These methodologies were implemented in a parallel crawler that captures the entire overlay topology of the system where each peer in the overlay is annotated with its available files. We have captured more than 50 snapshots of the Gnutella system that span over 1 year period. Using these snapshots, we conduct three types of analysis on available files: (1) Static analysis, (2) Topological analysis, and (3) dynamic analysis. Our results reveal several interesting properThis paper extends and supplants the earlier version of this paper presented at MMCN 2006 [1].

show abstract

Section: Capturing Unbiased Samplesmentioning

confidence: 99%

Section: Capturing Unbiased Samplesmentioning

confidence: 99%

Section: Capturing Unbiased Samplesmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Characterizing files in the modern Gnutella network

2007

View full text Add to dashboard Cite

show abstract

“…In addition, the fact that the random walk eventually visits all nodes an equal number of times, speeds up the algorithm's convergence by evenly providing return times to all nodes. The Metropolis-Hastings technique [23,33] is used to unbias the random walk. The node hosting the random walk selects a neighbor uniformly at random: the random walk is forwarded to the chosen neighbor with a probability depending on the degree of both nodes.…”

Section: Unbiased Random Walkmentioning

confidence: 99%

Second order centrality: Distributed assessment of nodes criticity in complex networks

Kermarrec

Merrer

Séricola

et al. 2011

Computer Communications

View full text Add to dashboard Cite

Gossip learning with linear models on fully distributed data

Ormándi

Hegedűs

Jelasity

2012

Concurrency and Computation

105

101

View full text Add to dashboard Cite

Machine learning over fully distributed data poses an important problem in peer-to-peer applications. In this model, we have one data record at each network node but without the possibility to move raw data because of privacy considerations. For example, user profiles, ratings, history, or sensor readings can represent this case. This problem is difficult because there is no possibility to learn local models; the system model offers almost no guarantee for reliability, yet the communication cost needs to be kept low. Here, we propose gossip learning, a generic approach that is based on multiple models taking random walks over the network in parallel, while applying an online learning algorithm to improve themselves, and getting combined via ensemble learning methods. We present an instantiation of this approach for the case of classification with linear models. Our main contribution is an ensemble learning method, which-through the continuous combination of the models in the network-implements a virtual weighted voting mechanism over an exponential number of models at practically no extra cost as compared with independent random walks. We prove the convergence of the method theoretically, and perform extensive experiments on benchmark data sets. Our experimental analysis demonstrates the performance and robustness of the proposed approach.EFFICIENT P2P ENSEMBLE LEARNING WITH LINEAR MODELS ON FULLY DISTRIBUTED DATA 557 and so on. Often, these personal data records are the most sensitive ones, so it is essential that we process them locally. At the same time, the learning algorithm has to be fully distributed because the usual approach of building local models and combining them is not applicable.Our goal here is to present algorithms for the case of fully distributed data. The design requirements specific to the P2P aspect are the following. First, the algorithm has to be extremely robust. Even in extreme failure scenarios, it should maintain a reasonable performance. Second, prediction should be possible at any time in a local manner; that is, all nodes should be able to perform high quality prediction immediately without any extra communication. Third, the algorithm has to have a low communication complexity, both in terms of the number of messages sent and the size of these messages as well. Privacy preservation is also one of our main goals; although in this study, we do not analyze this aspect explicitly.The gossip learning approach we propose involves models that perform a random walk in the P2P network, and that are updated each time they visit a node, using the local data record. There are as many models in the network as the number of nodes. Any online algorithm can be applied as a learning algorithm that is capable of updating models using a continuous stream of examples. Because models perform random walks, all nodes will experience a continuous stream of models passing through them. Apart from using these models for prediction directly, nodes can also combine them in various ways using ensemble learnin...

show abstract

On Unbiased Sampling for Unstructured Peer-to-Peer Networks

Cited by 200 publications

References 41 publications

Characterizing files in the modern Gnutella network

Characterizing files in the modern Gnutella network

Second order centrality: Distributed assessment of nodes criticity in complex networks

Gossip learning with linear models on fully distributed data

Contact Info

Product

Resources

About