Graph theory provides a variety of methods to characterize the way
vertices (peers) are connected, as static connectivity properties.
While structured P2P systems (such as DHTs) have well-defined
topologies based on maintaining certain mathematical properties, the
properties of unstructured P2P systems arise from a complex
interaction of peer churn, peer discovery, and neighbor selection.
Therefore, we have conducted a measurement study of a deployed P2P
system to study and characterize its graph properties.
Despite the fact that the overlay topology is such a fundamental
property, there have been few studies of the topologies of deployed
P2P systems. These studies are all at least three years old,
characterizing the Gnutella topology when it was just 3% of its
current size. Additionally, these studies do not capture significant
new topological features that have been introduced over the last few
year.
In , , we examine
static and dynamic properties of two-tier overlays (e.g., Gnutella,
eDonkey, FastTrack) using snapshots captured by Cruiser, using
Gnutella as an example. Our key results include:
- The peer degree is not power-law contrary to the result reported
in previous studies. In fact, we show how a power-law degree can be
the result of measurement error due to slow crawling.
- While not power-law, Gnutella is still a small world network. It
has short path lengths and more clustering than a comparably-sized
random graph.
- Long-lived ultrapeers form their own fairly well connected small
world graph, with increased clustering.
- The longer peers remain in the overlay, the greater the chance they
will be connected to other long-lived peers. There is an
onion-like biasing where peers are more likely to be connected to
older peers.
In , we extend our work by examining the geographic
distribution of peers and exploring long-term trends.
We have made some of topology snapshots available for the user of
other researchers here.