From: Danny Wyatt (danny@cs.washington.edu)
Date: Sun Nov 07 2004 - 22:54:29 PST
An Analysis of Internet Content Delivery Systems
Setfan Saroiu, Krishna Gummadi, Richard Dunn, Steve Gribble, Hank Levy
This paper presents 9 days worth of HTTP data observed between UW and
the internet. The analysis shows that P2P traffic, while having a lower
number of both requests and users, still occupies most of the network
traffic due to its far larger object size. I found two aspects of the
analysis particularly interesting.
First, their data supports a "some peers are more equal than others"
interpretation of P2P traffic. They observe that 600 (out of 281,026)
external P2P servers provide over a quarter of the incoming P2P bytes.
They also observe that UW P2P servers have an export/import ratio of
almost 7 to 1. P2P services allow users to choose the hosts from which
they want to download objects. Typically, these hosts are ranked by
observed bandwidth. My intuition suggests that hosts within UW have
more uplink bandwidth than home users (who typically have asymmetric
uplink/downlink speeds), and thus they appear more attractive as P2P
servers. Similarly, the top external P2P servers probably appear more
attractive for some criteria and thus are more popular. This
interpretation suggests that network inequities cause P2P services to
adopt a more traditional client-server behavior.
Second, they suggest that a local cache could provide all the benefit of
Akamai for content providers. They show that a small number of objects
account for most of the Akamai traffic. Additionally, though it goes
unmentioned in the text, Figure 11a shows that Akamai is doing its job
well: all traffic from UW is routed to a small handful of Akamai
servers. These two facts together suggest that Akamai is acting as a
local cache server, albeit one probably not tuned specifically to the
request signature of only UW. This conclusion is appealing, but not too
surprising. If there were adequate, reliable, trustworthy local caches,
then content providers would not need to hire Akamai. But because there
are none that fulfill all 3 of these needs---particularly
trustworthiness, or a certain liability---the providers must provide
these themselves, via Akamai.
This archive was generated by hypermail 2.1.6 : Sun Nov 07 2004 - 22:57:09 PST