From: Lillie Kittredge (kittredl@u.washington.edu)
Date: Sun Nov 07 2004 - 23:05:51 PST
Analyzing content delivery systems
This paper discusses a project to monitor and analyze all of the traffic
coming in and out of UW, concluding that a staggering amount of traffic is
on peer-to-peer systems.
The university is a convenient grounds for this sort of testing, since
we've got access to the gateway routers, and can see everything that's
coming and going. They found that www traffic was being utterly
overwhelmed by peer-to-peer traffic, which has larger objects and longer
downloads, and is also failing to scale well despite its inherent ability
to do so.
I found the discussion of dynamically generated content interesting - this
is getting more popular, but is harder to distribute efficiently since it
can't be cached. In general I agree with the authors that more and
smarter caching would make content delivery smoother.
What kills me about papers discussing peer-to-peer systems is the utter
glossing-over of issues of legality and propriety. The networking
community seems to assume that this is the wave of the future, it is what
the users want by huge margins, and so we should find a way to provide it.
I say, the majority of those huge files are pirated media, and, call me
outdated, but I believe that stealing is wrong. Other scientific and
engineering fields have maintained an awareness of ethics; what's wrong
with computer scientists that we don't? I would much prefer to see
literature discussing methods to discourage illegal media transfer.
Setting kazaa traffic at a lower priority and denying it bandwidth would
both alleviate the pressure it puts on WWW traffic and penalize this
mostly illegal practice.
That being said, if a non-evil use for peer-to-peer can be found, I agree
with the authors that caching could be a major improvement. What would be
more effective, were it easily possible, would be a way to prevent the
problem of uses becoming super-clients or servers, which defeats the
intention of peer-to-peer. Spreading load over the system (and preferably
doing so in a topologically sensitive way) would improve peer-to-peer
performance immensely. However, I suspect that the inherent greed of
peer-to-peer users prohibits this. The greed that makes users steal
pirated video is the same greed that makes users tend to avoid expending
bandwidth on allowing others to upload, and only stay connected long
enough to get what they want.
This archive was generated by hypermail 2.1.6 : Sun Nov 07 2004 - 23:05:52 PST