Evaluation of Cooperative Web Caching with Web Polygraph

1
Evaluation of Cooperative Web Caching with Web
Polygraph

• Ping Du and Jaspal Subhlok
• Department of Computer Science
• University of Houston
• presented at WCW 02, Aug 14

2
Web Cache Hierarchy
3
Motivation and Goals

• No practical methods to evaluate cache
  hierarchies under specific workload and network
  conditions
• Important for designing a caching solution
• Criteria for evaluation system
• Model reality well
• Applicable to different protocols structures
• Experiments should be repeatable
• Use both hit rate and user response time as
  metrics
• Solution based on Web Polygraph

4
Cache Evaluation with Web Polygraph
Polysrv
Polysrv
Polysrv

• Synthetic HTTP clients and servers on real
  machines on a LAN
• Workload parameterized by size, distribution,
  popularity, load and many others

Proxy Cache
Polyclt
Polyclt
Polyclt
5
Hierarchy Evaluation with Web Polygraph
Polysrv
Polysrv
Polysrv
Proxy Cache
Proxy Cache
Proxy Cache
Polyclt
Polyclt
Polyclt
6
Evaluation Framework

• Web Polygraph
• Reports throughput, response time, hit ratio etc.
  from clients viewpoint (but unaware of
  hierarchy)
• Dummynet
• Used to simulate networks of different
  capabilities by controlling bandwidth, latency
  and packet loss.
• Squid cache and Squeezer log analysis tool
• Captures cache cooperation info
• Modified to monitor specific polygraph phases
• Squeezer and Polygraph info has to be reconciled

7
Experimental Setup

• Experiments performed on different cache
  hierarchies of two, three four Squid caches.
• Hardware configuration of all Squid machines is
  the same (800MHz, 256MB, 4 30GB disks)
• Polygraph machines and caches on same 100Mbps
  switched ethernet network
• Balanced workload
• Cache fill-up phase not measured

8
List of Experiments

• Performance with different cache hierarchies
• Influence of network latency
• Influence of cache size
• Influence of the document sharing pattern
• One big cache compared to multiple caches
• Virtually unlimited experiment space with many
  parameters (e.g., request rate, public interest,
  cache, memory size etc.)

9
List of Cache Hierarchies
Cache
Client
2OY
3OY
2SY
Sibling-sibling
Parent-child
3SY
1ON-2OY
1ON-2SY
Same memory, disk per cache, fixed total request
rate, no network delay
2SY-1OY
2OY-1OY
1OY-2SY
10
Simulation Results - Different Hierarchies

• Improved hit ratio overcomes overheads of peering
• Parents appear less important than siblings

11
List of Experiments

• Performance with different cache hierarchies
• Influence of network latency
• 2 and 3 Squid caches independent or as siblings
• Network delay of 0 msecs, 40 msecs, or 80 msecs
  between caches
• Influence of cache size
• Influence of the document sharing pattern
• One big cache compared to multiple caches

12
Impact of Network Latency

• Hit ratio unaffected by latency
• Hit and Miss response times increase with latency
• Some increase in response time going from 0 to 40
  to 80 msec
• Cache cooperation is helpful even with modest
  network delay

13
Conclusion

• Web Polygraph based framework to evaluate
  cooperative caching
• Flexible
• Works on a real network
• Workload characteristics are easy to specify.
• Repeatable experiments
• Hit ratio and user response time based metrics
• Captures actual cooperation overheads

14
Future Work

• Make the toolset easily usable by the community
  currently a recipe type help available
• Evaluation of large hierarchies may need a
  combination of experimental and analytical
  methods
• More results from the performance of different
  kinds of hierarchies in different scenarios

15
Influence of Cache Size

• Two Squid caches, running isolated or as
  siblings.
• Various total disk cache size
• Same total memory cache size
• Same constant request rate
• No network latency between caches

16
Simulation Results - Cache Size

• Cooperative caches
• Higher hit miss response time
• Miss response time is stable.
• Increase in hit response time
• Fraction of memory to disk cache size
• Performance with increase of cache size
• Improve quickly
• Stabilizes gradually
• Benefits of cooperation increase.

17
Influence of the Document Sharing Pattern

• Two Squid caches, running isolated or as
  siblings.
• Various document sharing pattern
• Global URL space
• Public interest the percentage of all documents
  shared by Polygraph clients.
• Same total disk cache size
• Same total memory cache size
• Same constant request rate
• No network latency between caches

18
Simulation Results - Document Sharing Pattern

• Performance improves with public interest.
• Influence is mainly on remote hit.

19
Working Set Size
20
Performance of one big cache compared to multiple
caches

• One, two, three and four Squid caches
• Isolated or all siblings cache hierarchies
• Best effort workload
• Constant rate vs. best effort workload
• Used to get the best throughput
• Same total disk cache size
• Same total memory cache size
• No network latency between caches

21
Simulation Results - one big cache compared to
multiple caches

• One cache the worst throughput
• Two separate cache large improvement
• More separate cache declined performance
• All siblings hierarchy
• Improvement is more stable
• Levels off quickly
• Overheads of peering outweigh improved hit ratio
  eventually

22
Methodology - Phase Schedule

• Web Polygraph provides a scheme to customized
  desired workload pattern by phase schedule.

• Caches are in a stable state after Fill phase.
• Simulate daily Web traffic pattern in a short
  period.