From: Susumu Harada (harada@cs.washington.edu)
Date: Sun Nov 07 2004 - 19:40:12 PST
"An Analysis of Internet Content Delivery Systems"
S. Saroiu et al.
This paper analyzes the network bandwidth occupancy of different types of
content delivery systems in a university setting, and explores the
possibility of incorporating caching schemes to mitigate their impact on
the network. They looked at incoming and outgoing traffic for HTTP (WWW),
Akamai, Gnutella, and Kazaa applications over a period of nine days, and
made numerous observations regarding their network usage.
First, they noticed that the university had significantly more outbound
traffic than inbound, with peer to peer (P2P) traffic accounting for the
majority (Kazaa traffic was three times more than WWW). They also
observed that external Kazaa clients downloading content from the
university consumed nearly eight times more bandwidth than internal Kazaa
clients within the university. They also noted that compared to three
years prior to this study, transfer of multimedia files (video and audio)
increased by nearly a factor of four while HTML and image transfers nearly
halved. In summary, they found that P2P dominated WWW traffic, that P2P
transfers involved fewer clients but larger objects and longer transfers,
and that counter to the notion that P2P applications encourage
distribution of load, a very small number of external Kazaa peers provided
much of the incoming traffic to internal peers.
After making the above observations, the authors point out an issue of
concern regarding the scalability of P2P systems within an organization
with large population. They state that since "every peer in a P2P system
consumes bandwidth in both directions", the high bandwidth profile of
these applications can easily overwhelm the network's capacity. However,
their assumption about the symmetric nature of P2P transfer seems to
contradict their own observation, namely that inbound traffic was
significantly greater than outobund traffic. They also then conclude that
reverse caching to absorb outbound traffic, as well as a P2P cache, can
significantly reduce the burden on the network. However, they seem to
gloss over the cost implication of implementing such a cache, both from
economic point of view as well as performance point of view (the large
magnitude of the traffic that would need to be cached seem to require
significant storage and internal processing overhead). I think that the
idea of smart caching at the network boundaries of such large networks is
worth a closer inspection and analysis.
This archive was generated by hypermail 2.1.6 : Sun Nov 07 2004 - 19:40:13 PST