The Measured Performance of Content Distribution Networks

1
The Measured Performance of Content Distribution
Networks

• Kirk Johnson
• John Carr
• Mark Day
• Frans Kaashoek

2
Overview

• Background
• Procedure
• Results
• Discussion

3
What? And Why?

• Content distribution networks are coordinated
  caching systems.
• CDNs are a multi-million-dollar business already.
• Wed like to have a figure of merit for them.
• This talk doesnt have that figure of merit
• but it has some better data than just counting
  servers in networks

4
A DNS-redirecting CDN
DNS redirector
Client
http//example.com/foo
5
Client Actions

• R Resolve domain name
• F Fetch content
• Ordinary client use of CDN RF
• Instead of doing (RF) we do R then F
• This allows us to compare the server chosen to
  some other servers that could have been chosen,
  over a large number of fetches.

6
Procedure

• R Collect a set of servers by repeated DNS
  queries
• to a variety of name servers
• over a number of hours
• F Fetch a particular piece of content from each
  member of the set, measuring latency

7
Important Details

• Interleaved fetches
• Fetch1 at server1, fetch1 at server2, etc.
• Not fetch1 at server1, fetch2 at server1, etc.
• Unmeasured fetch before measured fetch
• Avoids cache misses
• Measure only HTTP fetch latency
• CDN not penalized for cost of DNS resolution

8
Where We Measured
9
Caveats

• Cant compare Akamai and DI directly
• The experiments fetched different-sized data
  (4672 vs. 3772 bytes)
• on different days
• This is not the definitive CDN-measuring paper
• Just some interesting indicators of whats out
  there

10
Akamai, location A
11
Looking at these graphs

• Note log plot of latency
• Gray line cumulative distribution at one server
• Red line cumulative distribution at all servers
• Blue line cumulative distribution at CDN

12
Cumulative Distribution

• Right way to look at this data
• Want to understand frequency and magnitude of bad
  choices
• Consistent vertical
• Fast to the left

13
Akamai, location A
14
Akamai, location B
15
Akamai, location C
16
Akamai observations

• Not optimal
• Does especially well in location B (Cambridge)
• Does only a fair job in location C (Boulder)

17
Digital Island, location X
18
Digital Island, location Y
19
Digital Island, location Z
20
Akamai observations

• Not optimal
• Does especially well in location B (Cambridge)
• Does only a fair job in location C (Boulder)

21
Akamai observations
Digital Island

• Not optimal
• Does especially well in location B (Cambridge)
• Does only a fair job in location C (Boulder)

Y
Z
22
Conclusions

• CDNs do add measurable value compared to random
  selection
• They arent perfect or optimal
• You can measure them yourself!
• Hard to know what explains performance
  differences
• Internal structures are secret