Neighbour selection strategies in BitTorrent-like Peer-to-Peer systems by rt3463df


									Neighbour selection
strategies in BitTorrent-
like Peer-to-Peer systems
L.G. Alex Sung, Herman Li
March 30, 2005
for CS856 Web Data Management
University of Waterloo

•   Introduction
•   Problem Statement
•   Proposed Categorization Schemes
•   Experimental Approach
•   Experiment Designs


• BitTorrent Highlights

        Problem Statement

• Goal: Explore the effect of two neighbour-selection
  strategies on the efficiency in content distribution for
  BitTorrent-like Peer-to-Peer systems
• Proposed strategies:
   – Neighbour selection by network capacity
   – Neighbour selection by locality
   – Preserve some degree of randomness
• What is BT-like? Incentive-built P2P systems with Tit-
  for-Tat exchange strategy. (central server not required)
• Why BT-like? We expect later unstructured P2P
  systems are BT-like. (eg. eXeem)
        Preserving Randomness

• Avoid power-law (Zipf) distribution of pieces:
   – Some pieces may be rare in one domain (capacity or
     locality), but popular in the other one
   – Even distribution of pieces increases the sustainability
• Randomness preserved by including some fraction of
  randomly chosen peers of a different domain

       Linear scales on both axes   Logarithmic scales on both axes   5
        Matching by Capacity

• Hypothesis: Matching peers according to capacity
  similarities improves efficiency due to the Tit-for-Tat
  exchange strategy
• When low ability peers are connected to high ability
   – get pieces when they are being optimistic unchoked
   – get choked again very quickly as they cannot offer a
     good exchanging rate

        Matching by Locality

• Hypothesis: Matching peers by locality:
   – Benefit from the lower network latency
   – Better utilization of bandwidth
• The topology of the overlay network better matches the
  underlying network
• In the case that the uploading capacity was not
  previously fully utilized:
   – maximize the uploading speed by exchanging with
     peers that are physically closer

          Experimental Approach

• Run experiments on Planet Lab nodes
• Planet Lab nodes experience similar network
  phenomenon as real BT users
• Select a set of Planet Lab nodes that is representative
  of the user population
• Population capacity and locality based on:
   – A public tracker log for “Beyond Good and Evil” from
     Dec 03 to Mar 04 [1]
     NL     Netherlands (Europe)   25269   46.90%
     US     United States          10250   19.02%   82.46%
     AU     Australia              6181    11.47%
     CA     Canada                 2730    5.07%

        Experiment Designs

• Experiment 1: Categorization by upload / download
  rate – sensitivity to randomness
   – System throughput vs randomness
• Experiment 2: Categorization by upload / download
  rate – sensitivity to number of categories
   – System throughput vs number of categories
• Experiment 3: Categorization by peer locality
   – System throughput vs randomness
• Experiment 4: Combination of improvement schemes
   – Categorization uses both capacity and locality
   – System throughput vs randomness

        Related work

• Anonymous BT with keyword search
   – eXeem (a commercial product w/ ads)
   – (IP is not shown directly in the GUI)

• Non-random peer set distribution
   – Based on content availability [2]


1. J.A. Pouwelse, P. Garbacki, D.H.J. Epema, H.J. Sips.
   The Bittorrent P2P File-sharing System:
   Measurements and Analysis. 4th International
   Workshop on Peer-to-Peer Systems (IPTPS'05), Feb
2. Simon G. M. Koo, C. S. George Lee, Karthik Kannan:
   A Genetic-Algorithm-Based Neighbor-Selection
   Strategy for Hybrid Peer-to-Peer Networks. In Proc. of
   the International Conference On Computer
   Communications and Networks (ICCCN 2004), IEEE
   2004, pages 469-474, October 2004.

To top