what-when-how
In Depth Tutorials and Information
We examine snapshots of the population of pieces in torrents, and the evolu-
tion of the piece population over several days, mostly during the early phases of
a torrent's lifetime. We find that the piece distributions are generally very nar-
row, and progress to more narrow distributions quickly in response to chang-
ing conditions. his shows that the downloading policy of BitTorrent is efective
from a piece distribution and evolution perspective, though we do find that some
enhancements are possible to achieve an ideal piece distribution, especially for
larger torrent.
Very little previous work has explored microscopic piece-level measurements of
BitTorrent. Niu and Li [27] attempted a theoretical evaluation of the
blockvariation
resulting from using network coding in a peer-to-peer system. hough focused on
network coding, their results should be applicable by setting the size of network
coding segments to 1, but they are only theoretical and do not match with our
experimental results.
he closest work to ours is from Legout et al. [23], who administrated a single
client and connected separately to 26 torrents of difering characteristics. heir
results thus reflect the piece availability only in peers their single client connected
to during the experiment, which may not be representative of the entire torrent,
nor does it offer global knowledge of the piece population. In contrast, our work
focuses on the global piece population by examining the number of copies of each
piece present in
every
peer in the torrent. We also follow the piece population over
time, to see how it evolves with the torrent. Some of this work has been previously
published [15].
6.2.1 Distribution of Piece Population
We now present and analyze our measurement results for the piece populations of
the torrents shown in Table 6.1. he results of some snapshots we took of the piece
population in some real Internet torrents are shown in section “Snapshots.” he
evolution of two real Internet torrents over long periods of time is shown in the
section “Evolution.” Finally, we ran several simulated torrents on PlanetLab, the
results of which are in section “Simulated PlanetLab Torrents.”
In order to make visual comparisons between different torrents, some normal-
ization of the data is needed. For all data, we normalize the number of copies (x-axis)
by the total number of downloaders, so that it varies from 0 to 1. If the populations
are all from the same torrent, the population size data (y-axis) will be normalized by
the number of pieces, so it will also vary from 0 to 1. However, this normalization
does not make sense when comparing different torrents' populations, as it leads to
a much smaller population when the number of pieces is larger. herefore, to facili-
tate the comparison of multiple torrents, they will be normalized so that the area
under their population distribution is 1. Since we are only interested in the width
of these distributions, this normalization should have no effect.
