Gnutella, FastTrack, and eDonkey use two-tier overlay topologies.

1
Daniel Stutzbach and Reza Rejaie University of
Oregon The Ion P2P Project http//mirage.cs.uoreg
on.edu/P2P
2. Two-Tier Topologies
1. Motivation

• Gnutella, FastTrack, and eDonkey use two-tier
  overlay topologies.
• Our initial study focuses on Gnutella.
• Top-level peers form the core overlay.
• Each leaf connects to a few top-level peers.

• Most of the large file sharing Peer-to-Peer (P2P)
  applications with millions of users are based on
  unstructured, two-tier overlay topologies.
• Characterizing the unstructured overlays of these
  applications is important for design and
  evaluation.
• Characterizing P2P overlays requires capturing
  accurate and fine-grained snapshots of the
  overlays.
• Snapshots (as graphs) are captured with a
  crawler, recording peers (as nodes) connections
  (as edges).
• Captured snapshots by a slow crawler can be
  distorted due to 1) dynamic changes in the
  overlay during a crawl, 2) peers unreachable by
  the crawler.
• Previous studies are outdated, used slow crawlers
  (1 or 2 hours), and did not examine the accuracy
  of their captured snapshots.

3. Approach

• We developed a parallel and tunable crawler,
  Cruiser.
• Cruiser increases crawling speed by
• Using a master-slave architecture
• Crawling many peers in parallel
• Leveraging the two-tier topology
• Cruiser captures an accurate Gnutella snapshot
  with 1-million peers in around 7 minutes (140k
  peers/min).

4. Results
Top-to-Leaf Degree Distribution
Leaf-to-Top Degree Distribution
Inaccuracy of Slow Crawling
Top-to-Top Degree Distribution
Fig. 4
Fig. 3
Fig. 1
Fig. 2

• Peer degree is fairly homogeneous, not power-law.
• Most top-level peers have a degree around 30
  (Fig. 1).
• Under 30, degree is nearly uniformly distributed
  (Fig. 1).
• A power-law was reported by previous studies.
• Our prior work shows that slow crawling can
  erroneously lead to a power-law degree
  distribution (Fig 2).

• Degree distributions between tiers are also
  fairly homogenous.
• The number of leaves per top-level peer is
  similar. Version differences cause different
  spikes (Fig. 3).
• Most leaf peers have a very low degree, but a
  small number have a high degree (Fig. 4)

Small World Properties
Path Length Distribution
Graph Path Lengths Lengths of Random Clustering Coefficient CC of Random
Modern Gnutella 4.174.23 3.75 0.018 0.00038
Older Gnutella 3.304.42 3.66 0.02 0.002
Movie Actors 3.65 2.99 0.79 0.00027
Power Grid 18.7 12.4 0.08 0.005
Fig. 5
Fig. 6

• Despite exponential growth in size, overlay path
  lengths in Gnutella are very short (Fig. 5 and
  Fig. 6).
• 60 of top-level paths are exactly four hops in
  length.
• 99.5 of top-level paths are five hops or less.
• Leaf-to-leaf paths are 1 or 2 hops longer, on
  average.

• Gnutella is not power-law, but is still a small
  world.
• Path lengths are close to same-size random
  graphs.
• The top-level overlay is not tightly clustered
  (0.018).
• However, it is much more clustered than same-size
  random graphs (0.018 gtgt 0.00038).

5. Ongoing Work

• Characterizing file distribution and query
  workload
• Characterizing Kademlia-based DHTs
• Examination of performance bottlenecks in
  BitTorrent

• Characterizing modeling the dynamics of overlay
  topologies 1) Peer churn, 2) Edge churn
• Developing an overlay topology generator for
  simulation