Web Server Workload Characterization The Search for Invariants

1
Web Server Workload Characterization The Search
for Invariants

• John Oleszkiewicz
• CSE 807
• January 14th, 2003

2
Problem

• WWW is the biggest contributor to packet and byte
  traffic on the Internet (1996)
• WWW is growing at an astonishing rate
• 1992 74 MB of traffic
• 1996 3.2 TB of traffic
• We want to reduce traffic and response latency
• We need a workload characterization

3
Data

• Websites
• University of Waterloo CS department
• University of Calgary CS department
• University of Saskatchewan University website
• NASA's Kennedy Space Center
• ClarkNet (ISP)
• National Center for Supercomputing Applications

4
Data

• Ranges
• Traffic
• 776 requests per day
• 355,787 requests per day
• Time
• 1 week
• 1 year

5
Data

• File access log
• Name of client
• Time of request
• Name of document
• Server's response
• Number of bytes transferred

6
Invariants

• What observations apply across all data sets
  studied?

7
Invariant 1

• Success rate for server file lookups is 88
• Success is one of four possibilities
• Successful
• Not Modified
• Found
• Unsuccessful
• Not modified came in second at 8

8
Invariant 2

• HTML and image files account for 90-100 of
  requests
• Documents were classed by file extension
• Other file types include sound and video
• Result is consistent with other studies

9
Invariant 3

• Mean transfer size is less than 21 KB
• Most files are small

10
Invariant 4

• Among all server requests, less than 3 of the
  requests are for separate files
• Range 0.3 to 2.1

11
Invariant 5

• Approximately 1/3 of the files accessed are
  accessed only once in the log

12
Invariant 6

• File size distribution is Pareto distribution
• Most files are between 100 and 100,000 bytes
• 10 of files are larger than 100,000 bytes.
• Consistent with other studies and literature.

13
Invariant 7

• 10 of files accessed account for
• 90 of server requests
• 90 of bytes transferred
• Example sound and video account for
• 0.1 to 1.2 of requests
• 0.2 to 30.8 of data

14
Invariant 8

• File inter-reference times are exponentially
  distributed and independent
• Inter-reference time between accesses
• Mean depends on workload

15
Invariant 9

• Remote sites account for
• More than 70 of accesses
• More than 60 of bytes transferred

16
Invariant 10

• Web servers are accessed by thousands of domains
• 10 of domains account for more than 75 of usage

17
Other findings

• No temporal locality
• Self-similarity on some workloads

18
Recommendations

• Server-side caching
• Large number of references to a small number of
  files
• Large number of bytes transferred from a small
  number of files

19
Caching Strategies

• Cache large number of small files
• Allows more requests to be handled
• 76 of all file references are less than 10KB in
  size
• Reduces CPU load
• Cache small number of large files
• Reduces bytes transferred by server
• Only 26 of byte traffic is generated by files
  smaller than 10Kb
• Reduces disk load

20
Other Issues

• One-time file referencing
• Means one-third of server cache can be cluttered
  with useless files
• No temporal locality
• LFU cache replacement policy is attractive
• LRU cache replacement policy is not

21
Conclusion

• Server-side caching can yield significant
  performance increases
• Which caching scheme is used depends on system
  bottlenecks (disk, CPU)

22
THE END

• Thanks for listening!