Adaptive Schemes for Distributed Web Caching

1
Adaptive Schemes for Distributed Web Caching

• Spiridon Bakiras
• Thanasis Loukopoulos
• Dimitris Papadias
• Ishfaq Ahmad

2
What is a Proxy Server?

• Computer that answers requests for a connection,
  file, or other resource available on a different
  server
• Common proxy application is a caching Web proxy

3
Proxy Server
4
What is Proxy Caching?

• Cached a local copy of recently requested web
  pages locally
• Goal is to reduce latency

5
Cache Sharing Architectures

• Architecture for sharing the cache
• Hierarchical
• Distributed
• Hybrid

6
Hierarchical Caching

• Proxy server sharing caches by parent-child
  relationship
• Problems
• Possible slow links
• Upper level nodes my become overloaded

7
Distributed Caching

• Relationships without an upper level
• Static or dynamic

8
Squid Web Proxy Software

• Most popular proxy software (FREE)
• Hybrid architecture both hierarchical
  distributed
• Uses Internet Cache Protocol (ICP)

9
Squid

• If cache miss occurs
• First it broadcasts to its neighbors
• Takes the first neighbor to respond
• If no response, forward to parent proxy or the
  web server itself

10
Cache Digests

• Proxies exchange their directory information
• Proxy check the digest local requests only the
  neighbor that have the cached page.

11
Choosing Neighbors

• Neighbors should be
• Near by in terms of network latency
• Similar in surfing behaviors

12
Problems Paper Addresses

• Possible to improve the performance of current
  distributed web caching schemes using neighbor
  reconfiguration.
• And can we tackle the resulting problem as a
  second level caching.

13
Second Level Caching

• Unidirectional (or asymmetric) caching
• Dynamically
• Keeps adds best neighbors
• Evicts worst
• Bidirectional (symmetrical) caching
• Handle both incoming outgoing digests

14
Unidirectional Caching

• Goal dynamically update its neighborhood list
  in order to maximize the hits from other proxies.
• Replace least beneficial neighbors with ones that
  may provide more hits.
• Techniques
• Least Recently Used - LRU
• Least Frequently Used - LFU

15
Unidirectional LRU

• If a cache miss occurs locally
• Broadcast sent to all neighbors
• If multiple neighbors can serve the request, the
  closest in latency is chosen
• If none of the neighbors can serve the request,
  the exploration process starts
• If proxy returns a request from exploration
  process, it become the newest neighbor.
• The neighbor with the oldest response is released

16
Unidirectional LRU Problem

• Network overhead due frequent reconfiguration of
  neighbors exchange of digests
• Statistical good neighbors might have short
  unproductive period, even though it would be good
  for the long run

17
Unidirectional LFU

• Statistical information is keep on hits from all
  the neighbors.
• A neighbor is replaced when a statistically
  better proxy is found

18
Bidirectional caching

• Cache digest information is exchanged with each
  neighbor.

19
Experiments

• Unidirectional LRU vs. static methods
• Better hit ratio then static with same number of
  neighbors
• Unidirectional LFU vs. LRU
• LFU has less network overhead then LRU
• Bidirectional
• Less overhead from digest exchange

20
Experiment Summary

• Unidirectional strategies achieve the best
  performance
• Neighborhood Hits
• Traffic Overhead
• Unidirectional LFU is considerably lower amount
  of traffic overhead then LRU due to reduce
  neighbor list reconfiguration.

21
Experiment Summary (contd)

• Bidirectional strategies are best with load
  balancing.
• Bidirectional LFU get the best results.

22
Conclusion

• Both unidirectional LFU LRU achieve higher hit
  ratios compared to a static configuration
• Second level caching could be used to solve
  problems of latency, recency, frequency of
  requests, etc.

23
thank you