Collaborative Web Caching Based on Proxy Affinities

1
Collaborative Web Caching Based on Proxy
Affinities

• Jiong Yang, Wei Wang in T. J.Watson Research
  Center
• Richard Muntz in Computer Science Department of
  UCLA
• Proceedings of the international conference on
  International Conference on Measurements and
  modeling of computer systems, 2000, Pages 78 - 89

2
Outline

• 1. Introduction
• 2. Related Work
• 3. Objective Model
• 4.Page Cluster
• 5.Information Group Maintenance
• 6.Web Page Retrieval
• 7.Experience Result
• 8.Estimation of Information Group Size

3
1. Introduction

• Recent research to improving internet performance
  into three categories
• --server load balancing
• --intra-net collaborative caching
• (summary cache)
• --inter-net collaborative caching
• 1. Nearby proxy faster than distant server
• 2. A proxy with up-to-date page could
  serve as a server
• Drawback Burst network traffic

4
Introduction(cont.)

• Each request full into three categories
• --The proxy locally cached the up-to-date
  version of web page
• --The up-to-date web page exists on nearby
  proxies
• --The requested web page has to obtained from
  the content server

5
Introduction(cont.)

• How to discovery which proxy has cached web page?
• --pull (more response time)
• --push (more communicating messages)
• In this paper
• --dynamic distributed collaborative caching
  infrastructure
• --information group ( web page clusters )
  proxy profile( list of URLs )
• --GOALmessges (among proxies update)
• maintain cache hit rate
  latency

6
2. Related Work

• Caching in Harvest
• --caches organized in a hierarchy
• Adaptive Web Caching
• --self-organizing to form a tight mesh
• Summary Cache
• --Each proxy keep a summary (using cache
  sharing protocol)

7
Related Work(cont.)

• Web Caching Based on Dynamic Access Patterns
• --A local caching algorithm flexibly adapts
  its parameters
• Server Volumes and Proxy Filters
• --piggyback

8
3. Objective Model

• ? local cache hit ratio
• w remote cache hit ratio
• Local-cost
• Remote-cost
• Server-cost
• Locating-cost find where the cache is
• Push-cost incurred multicast by changes
• Serach-cost Push-costLocating-cost
• Cost

9
Objective Model(cont.)

• Imax , I min the number of proxies in a
  collaboration
• m cache hit ratio?gt search cost

10
4. Page Cluster

• Frequency this web page / total pages
  
• threshold(ß)
• Grouping web pages into clusters
• --Each proxy sends its profile to a central
  site S
• --An optimal or near optimal partition of
  frequency accessed web pages is generated

11
Page Cluster(cont.)
12
Page Cluster(cont.)
13
Page Cluster(cont.)
The number of cluster
We need additional data structure
14
Page Cluster(cont.)

• The action on a page
• --move to another cluster
• --replicate in another cluster
• --remove replica from this cluster

15
Page Cluster(cont.)
16
Page Cluster(cont.)

• Choose a server to be the coordinator of
  information group
• The content of all page clusters and their
  coordinators are broadcast to all proxies

17
5. Information Group Maintenance

• Each information group is associated with one
  page cluster.
• A proxy join a information group which has
  maximum pages in it.Find another ..until the
  proxy joins the information groups for all web
  pages on its profile.
• Local reorganization

18
Information Group Maintenance(cont.)

• A proxy wants to join an information group
• --send a message to coordinator of information
  
• group
• --send back the list of the members
• --the new proxy send the intersection of its
  cache content to all member in this information
  group
• A proxy wants to withdraw from a information
  group
• --multicast to all member
• If a proxys cache for a page cluster changes by
  more than 10 as the threshold ,multicast to all
  member( the lowest priority )

19
6. Web Page Retrieval
20
7. Experience Result
21
Experience Result(cont.)
22
Experience Result(cont.)
23
Experience Result(cont.)
24
Experience Result(cont.)
25
8. Estimation of Information Group Size

• Cost

26
9. Conclusion

• Dynamic adaptable structure
• Good scalability
• Maintain a high hit ratio and less latency and
  less message