CPSCox: A survival analysis model of peer behavior in large scale DHT system

  • Authors:
  • Daochao Huang;Fuhong Lin;Lei Wu;Hongke Zhang

  • Affiliations:
  • National Engineering Lab for Next Generation Internet Interconnection Devices, School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China;National Engineering Lab for Next Generation Internet Interconnection Devices, School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China;National Key Laboratory of Universal Wireless Networks, School of Information and Communications Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, China;National Engineering Lab for Next Generation Internet Interconnection Devices, School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China

  • Venue:
  • Computer Communications
  • Year:
  • 2012

Quantified Score

Hi-index 0.24

Visualization

Abstract

The peer behavior of P2P network has become a major concern and attracted significant attention in recent years. Most existing peer behavior research primarily focuses on only some specific properties of peers or requires the knowledge of detailed parameter values, which makes their analytical models not adoptable for large scale and dynamic Distributed Hash Table (DHT) system. In this paper, we propose a general recurrent events modeling in which three major types of peer behavior in DHT systems, session length, inter-session length and remaining uptime are considered. This model, called CPSCox, combines the counting process and stratified Cox proportional hazards method to explicitly reveal critical risk factors that influence the peer behavior and find out the distribution of session length and inter-session length of peers. Real dataset gathered from realistic KAD networks were employed to verify our model. Evaluation results illustrated that the model is able to obtain adequately reliable estimates of the regression coefficients for session length and inter-session length even though the baseline hazard or survival is not specified. The effective of predicting remaining uptime in large scale KAD-like DHT systems is validated as well. Being a semi-parametric method, CPSCox can closely approximate to correct parametric models.