An Analysis of Internet Content Delivery Systems

1
An Analysis of Internet Content Delivery Systems

• Stefan Saroiu, Krishna P. Gommadi, Richard J.
  Dunn, Steven D. Gribble, and Henry M. Levy
• Proceedings of the 5th Symposium on Operating
  Systems Design and Implementation
• December 2002

2
Outline

• Goals of Paper
• Overview of Content Delivery Systems
• Experimental Methodology
• Results
• Caching
• Conclusions

3
Goals

• Quantify the increasing importance of novel
  content delivery systems
• Characterize the behavior of these systems from
  the perspectives of clients, objects, and servers
• Derive implications for caching in these systems

4
Content Delivery Systems

• HTTP Web Traffic
• Content Delivery Networks
• Akamai
• Peer-to-peer file sharing networks
• Gnutella
• Kazaa

5
HTTP Traffic

• Clients request objects from web servers using
  HTTP
• Most web objects are small, 5-10KB.
• Web object requests follow a Zipf-like
  distribution
• Caching
• Cache hit rate increases logarithmically with
  client population
• Impossible for dynamic content

6
Zipf Distribution Compared with Sun Log Data
7
Content Delivery Networks (CDNs)

• Dedicated collections of servers that are
  geographically distributed
• Provide static content, e.g. images, streaming
  video
• Allows user to access replica of content that is
  close
• Replica location done via DNS interposition or
  URL rewriting at origin servers
• Redirection adds overhead
• Reduces average download response time

8
Peer-to-Peer Systems

• Peers form a distributed system to exchange
  content
• Batch-style downloads
• Most peers have low-availability and limited
  network capacity
• Files transferred via direct connection between
  peers

9
Experiment Methodology

• Use passive network monitoring to collect trace
  of TCP traffic between University of Washington
  (UW) to rest of Internet
• Collected 9 days of data, over 20 TB

10
Some Interesting Observations

• UW is an HTTP content provider
• Exported 16.65 TB. Imported 3.44 TB
• Bandwidth consumption (inout)
• .2 Akamai
• 6.04 Gnutella
• 14.3 WWW
• 36.9 Kazaa
• Rest is other TCP protocols mail, streaming
  video/audio, etc.

11
Some More Interesting Observations

• Compared to 1999 study
• HTML traffic has decreased 43
• GIF/JPG traffic has decreased 59
• AVI/MPG traffic increased nearly 400
• MP3 traffic increased nearly 300

12
Objects

• Median P2P object size is 4MB.
• Median Web object is 2KB
• 5 of Kazaa objects are over 100MB
• Top 1 of Kazaa objects account for 50 of bytes
  transferred
• For Web, top 1 account for 16 of bytes
  transferred

13
Clients

• For both Web and Kazaa, small number of clients
  account for large portion of traffic
• In Web, top 200 clients (0.5 of the population)
  account for 13 of the traffic
• In Kazaa, top 200 clients (4 of the population)
  account for 50 of the traffic

14
Servers

• Would expect server load for Kazaa to be much
  more distributed than for WWW
• This is not the case
• Top 500 external Web servers provide 22 of the
  bytes
• Top 500 external Kazaa servers provide 10 of the
  bytes

15
Scalability

• With respect to bandwidth cost adding another
  450 Kazaa clients would be equivalent to doubling
  the web client population (from 40,000 to 80,000)

16
CDN Caching

• Do CDNs provide any performance benefits over
  local proxy cache?
• If Akamai traffic were directed to proxy cache
  instead
• 88 ideal object hit rate (all objects cacheable)
• 50 practical hit rate
• Conclusion Widely deployed proxy caches reduce
  need for separate CDNs

17
P2P Caching

• Inbound cache byte hit rate 35
• Outbound cache byte hit rate 85
• Hit rate increases with client population
• 1,000 clients 40 hit rate
• 500,000 clients 85 hit rate
• Conclusion Reverse P2P cache saves the most
  bandwidth

18
Conclusions

• P2P traffic accounts for majority of HTTP bytes
  transferred
• P2P objects are significantly larger than Web
  objects
• Small number of large objects account for a large
  percentage of P2P traffic
• Small number of clients and servers responsible
  for majority of P2P traffic
• P2P traffic creates significant bandwidth load