Review: Gummadi et al., ...Analysis of a Peer-to-Peer...

From: Steve Arnold (stevearn_at_microsoft.com)
Date: Sun Mar 07 2004 - 22:27:11 PST

  • Next message: Cliff Schmidt: "Gummandi et al. "Measurement Modeling, and Analysis of a Peer-to-Peer File-Sharing Workload""

    In this paper, the researchers aren't designing anything (well there's a
    few ideas in there), but rather they are taking a look at peer-to-peer
    file sharing workloads. (The scope is quite specific: the use of Kazaa
    at the UW.) They started their analysis by collecting a 200 day trace of
    internet packet data, where the Kazaa requests were filtered on.

     

    First, they made several key observations based on the data. Kazaa users
    will wait a long time just to download files (in contract to web users).
    As users are on the system longer, they tend to slow down in their
    requests. Most requests are for smaller files. However, given this, "the
    majority of bytes transferred are due to the largest objects (video
    files)." Perhaps the most key observation is that Kazaa objects are
    typically "fetch-at-most-once." In addition, popular objects are usually
    short-lived, and, at the same time, most requests are for older objects.

     

    The next section argues that "Kazaa is not Zipf." There have been
    previous studies that might indicate this (esp. for the web in general).
    However, it is not the case for Kazaa mostly because the objects are
    immutable and they are fetch-at-most-once. They show this by modifying
    some parameters in a simulation. This does not prove their point, but it
    presents a reasonable argument.

     

    Lastly, they explore ways to reduce this bandwidth demand (over half of
    internet traffic - it would be interesting to know what effect, if any,
    that legal pursuits of the recording industry have had). The natural way
    is place caches in certain networks. However, many institutions may not
    be willing to do this for legal concerns. Their suggestion is
    redirection. When a request goes out for an object that exists within
    the sub-network, then the packet is re-directed to that client. They
    found that they could achieve between a 37% and 63% hit rate, enough to
    reduce external peer-to-peer traffic by over 80%.


  • Next message: Cliff Schmidt: "Gummandi et al. "Measurement Modeling, and Analysis of a Peer-to-Peer File-Sharing Workload""

    This archive was generated by hypermail 2.1.6 : Sun Mar 07 2004 - 22:27:36 PST