From: Andrew Putnam (aputnam@cs.washington.edu)
Date: Fri Nov 05 2004 - 17:06:37 PST
An Analysis of Internet Content Delivery Systems
Stefan Saroiu, et al.
Summary: Traffic patterns at the University of Washington are analyzed
and broken down into four different categories: HTTP web traffic,
Akamai CDN traffic, peer-to-peer traffic, and other. These results show
that peer-to-peer traffic has substantially different characteristics
than standard web traffic and accounts for a substantial percentage of
Internet traffic. The authors conclude that peer-to-peer networks
cannot scale on the current Internet architecture and that current
caching mechanisms are inadequate.
The Internet architecture has supported a wide variety of different
applications, but most of those applications had similar traffic
patterns: small, short, interactive transfers. These patterns allowed
Internet architects to make assumptions about the mechanisms necessary
for providing quick, reliable content delivery. The advent of
successful P2P networks changed the dynamics of network traffic and
challenges the assumptions of typical network traffic.
The authors find that the P2P networks tend to have much larger files
and much longer duration transfers than HTTP web traffic. The most
serious consequence of this is that a very small number of users can
account for a disproportionate amount of network bandwidth. A mere 200
users out over 65000 accounted for over 20% of the total network
traffic during the study. This suggests that P2P networks simply cannot
scale like HTTP web traffic since the network demands of each P2P user
are orders of magnitude greater than web users.
One way in which the Internet is not designed for P2P traffic is there
is the assumption that users are waiting on the network, so the network
needs to deliver packets as quickly as possible. With P2P networks,
users are willing to wait long periods of time for file transfers. They
tend to start file transfers late a night and are willing for those
file transfers to continue for long duration. This traffic travels at
the same priority as interactive traffic, which reduces the response
time for users who truly do care about network latency. If the Internet
had a priority system for packets, P2P traffic could safely run at a
lower priority without degrading the effectiveness of the P2P network.
Another way in which the Internet is not designed for P2P traffic is
the network hardware architecture. The Internet is set up in a
hierarchy of network providers. Servers tend to be closer to the top of
the hierarchy where bandwidth is plentiful. Clients tend to be near the
bottom of the hierarchy since they have much smaller bandwidth needs.
This asymmetric architecture allows clients to be as close as possible
to the servers, and the servers to be able to serve many different
clients. In P2P networks, clients become servers as well as clients.
This makes clients from distant parts of the network go through
numerous other networks and into a low bandwidth network in order to
get files. The low bandwidth networks become bottlenecks for both the
clients and the P2P servers.
One particularly interesting conclusion the authors came to is that the
University of Washington could alleviate network congestion by adding a
cache, but not in the typical sense of caching the most frequently
accessed data to make it easily available to students at UW. Instead,
it makes more sense to cache frequently accessed data within the UW
network on a cache accessible from outside UW. This would keep outside
users from using UW network bandwidth while connecting to UW servers.
If P2P systems regain their prominence, the Internet architecture will
have to evolve to handle the different traffic characteristics, or else
the P2P services will not scale, and will degrade the performance of
traditional Internet applications.
This archive was generated by hypermail 2.1.6 : Fri Nov 05 2004 - 17:06:36 PST