Sampling and Peer-to-Peer Networks: Is What You See Really What You've Got?

• Activity Number: 4
• Type: Invited
• Date/Time: Sunday, July 29, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Eric D Kolaczyk*+ and Qi Ding and Mark Crovella
• Companies: Boston University and Boston University and Boston University
• Keywords: Network ; Sampling ; Internet ; Stochastic block model

Abstract

Peer-to-peer (P2P) networks in the Internet - such as derive from the Skype, Bittorrent, and Gnutella apps - are self-organized, temporary communities involving a decentralized exchange of information between actors within these communities. Monitoring such networks is important for a range of issues associated with network architecture, search, and security. It has been remarked that in at least one such class of P2P networks the underlying protocol itself can be expected to induce a particular stratified structure, where stratification is associated with differences in information bandwidth of participant nodes. In this work we examine the effect of standard Internet packet sampling on our ability to accurately infer these networks and their characteristics from measurements. Using a stochastic block model formulation to represent the stratified P2P network, we show that the interaction of network structure and packet sampling leads to our seeing fewer nodes, differentially with respect to bandwidth, while, at the same time, more cohesive subgraphs. We discuss statistical procedures to correct the underlying sampling bias for estimating certain key network quantities.