Performance Evaluation of Redirection Schemes in Content Distribution Networks

1
Performance Evaluation of Redirection Schemes
inContent Distribution Networks

• Jussi Kangasharju, Keith W. Ross
• Institut Eurecom

Jim W. Roberts France Telecom RD
2
Outline

• Content Distribution Networks
• Redirection Schemes
• Simulations
• Experiments
• Discussion
• Future Work
• Conclusion

3
Content Distribution Networks

• Content distribution on the Web is changing
• Content placed on content servers
• Clients redirected to nearby content servers
• Better performance for clients
• Load balancing on content servers
• Better fault tolerance

4
Redirection Schemes

• Full redirection
• All client requests directed to content servers
• Advantage Simple
• All content servers must be able to serve any
  content -gt full mirrors or proxies
• Selective redirection
• Only some requests redirected
• Need to select what content to serve from content
  server

5
DNS Redirection

• Currently CDNs use DNS redirection
• Two strategies
• CDN controls the DNS resource records of the
  origin server
• References to objects are changed to point to a
  content server
• First strategy Full redirection
• Second strategy Selective redirection

6
Simulation Model

• Use the NS network simulator
• Simple model Client, Server, and Link
• Vary RTT, bandwidth, and loss rate on link
• Client requests HTML file, then requests images
• Look at user-perceived performance (total time)
• Does not account for server load (yet)

7
Effects of Bandwidth

• After 1.5 Mbps, download time mostly based on RTT
• Study only RTT and loss rate

8
Different Clients

• Two different clients Pipelining and parallel
• Assumption Parallel client cannot retrieve all
  objects with one set of connections
• Baseline Retrieve all from origin server
• RTTs 10, 20, 60, 100, 120, and 160 ms
• Does not include DNS lookup delay
• Assume parallel connections to be independent

9
Test Files

• 100 pages from Hot100.com
• Typical homepage 20 KB 50 KB

10
Loss Free Network, Baseline

• Big gains possible, up to 90

11
Loss Free Network, Switch

• Parallel connections, Medium page
• Gains limited, maximum 30

12
Loss in Network, Baseline

• Big gains still possible, up to 85
• Not as good as no-loss case (slow-start)

13
Loss in Network, Switch

• Parallel connections, Medium page
• Maximum gains 20 (slow-start, loss of SYN)

14
Experiments

• Repeat simulated experiment on Internet
• Exclude DNS lookups
• Results confirm results from simulations

15
Discussion

• Switching servers limits performance
• Either new server is not fast enough or the
  client should have used it for all the objects
• Easy with full redirection, can be done with
  selective redirection, but have to be careful
• Persistent connections
• No server load -gt switching degraded
• No redirection cost -gt switching improved

16
Future Work

• Improve simulation topology and model
• Take into account server load in simulations
• Include cost of redirection
• Perform more extensive experiments
• Evaluate the server selection schemes of existing
  CDNs

17
Conclusion

• Evaluate effects of redirection on user-perceived
  performance
• Both simulations and experiments
• Switching servers during download of a page hurts
  performance
• Need to replicate Web pages with images to
  guarantee best performance