Web%20Cache%20Replacement%20Policies:%20Properties,%20Limitations%20and%20Implications

1
Web Cache Replacement Policies Properties,
Limitations and Implications

• Fabrício Benevenuto, Fernando Duarte, Virgílio
  Almeida, Jussara Almeida

Computer Science Department Federal University of
Minas Gerais Brazil
2
Summary

• Introduction to Web caching
• Motivations and goals
• Evaluation methodology
• Performance metrics
• Workload description
• Caching system simulator
• Experimental results
• Conclusions and future work

3
Web Caching

• Dramatic growth of the WWW in terms of content,
  users, servers and complexity
• Web caching is a common strategy used to
• reduce the traffic over Internet
• increase server scalability
• diminish the latency in the network
• Use of caching by the deployment of Web Proxies

4
Web Caching

• Web proxies can be seen as intermediaries of the
  traffic between the HTTP clients and servers
• Nowadays the Web has a hierarchical topology

5
Web Caching

• Cache replacement is one of the issues that a
  proxy should be able to manage
• As the cache has finite size, when it is full,
  how does a proxy choose a page to remove from its
  cache?
• A lot of research has been done to address this
  question and several cache replacement policies
  can be found in the literature
• Key questions
• Is the design of new cache replacement policies
  needed?
• What are the properties that new policies should
  take advantage of to improve a caching system?

6
Goals

• Investigate how much a new caching policy could
  improve cache system performance
• Explore the main causes of periods of poor and
  high performance in caching systems

7
Evaluation Methodology

• Evaluation of different metrics over time
• Hit Ratio
• Percentage of first-timers
• Maximum improvement
• Entropy
• Time intervals of 1, 10 and 100 minutes
• Use of real workloads

8
Performance Metric Hit Ratio

• Hit ratio is the percentage of requests satisfied
  by the cache
• It is most general metric used to evaluate the
  effectiveness of a caching policy
• Measuring hit ratio over time to detect periods
  of variations of performance

9
Performance Metric Percentage of First-Timers

• First-timer is the first request for an object of
  the trace.

• Caching policies cannot satisfy first-timers
• the first-timer has never been requested in the
  past

10
Performance Metric Maximum Improvement

• The maximum improvement MI is defined as

• Maximum improvement over LRU

• We evaluate the maximum hit ratio a new caching
  policy can improve over the simple LRU policy

11
Performance Metric Entropy

• Taking n distinct objects with probability pi of
  occurrence, the entropy H(X) of a request stream
  is calculated as

• Entropy measures the concentration of popularity
  of a request stream
• The higher the value of the entropy, the lower
  the concentration of popularity
• Caching policies should keep objects with high
  probability of being referenced in the near
  future

12
Performance Metric Entropy

• Entropy depends on the number of distinct objects

• Use of the normalized entropy HN

• Investigate the influence of popularity on
  caching performance

13
Experiment Setup

• Real traces from proxy caches located at two
  points of the Web topology
• Closer to clients
• Federal University of Minas Gerais (UFMG)
• Closer to servers
• National Laboratory for Applied Network
  Research (NLANR)
• Cache Size 10 of the number of distinct objects
• Replacement caching policy Simple LRU

14
Workload Description
Name University 1 University 2 NLANR 1 NLANR 2
start date 01-10-2004 01-12-2004 01-18-2005 01-20-2005
days 2 10 2 11
requests 1,004,747 3,459,549 1,207,075 3,427,391
distinct objects 299,367 623,164 891,906 2,350,215
normalized entropy 0.8532 0.8268 0.9482 0.9329

• Traces used
• Cache warming University 1, NLANR 1
• Performance evaluation University 2, NLANR 2
• Higher concentration of popularity on university
  traces (lower entropy)
• Larger fraction of different objects in the NLANR
  traces, what diminish significantly the caching
  performance

15
Experimental Results Hit Ratio
proxy closer to clients
proxy closer to servers

• Higher hit ratio for University trace
• Strong variation along the time
• What are the factors that causes the variations
  on hit ratio?

16
Experimental Results Percentage of First-Timers
proxy closer to clients
proxy closer to servers

• Smaller of first-timers at the proxy closer to
  clients
• Correlation coefficient between hit ratio and the
  percentage of first-timers
• -0.857 for the NLANR and -0.962 for the
  university
• Caching policies cannot satisfy first-timers, the
  most important factor for poor and good
  performance in the analyzed traces

17
Experimental Results Entropy
proxy closer to clients
proxy closer to servers

• Proxy closer to clients lower entropy ? higher
  concentration of popularity
• LRU policy does not take advantage of all
  locality of reference
• Correlation coefficient between hit ratio and
  entropy
• -0.787 for the NLANR and -0.453 for the
  university
• If we had a caching policy able to filter all the
  locality (entropy 1), how much could hit ratio
  be improved?

18
Experimental Results Maximum Improvement
proxy closer to clients
proxy closer to servers

• The hit ratio cannot be significantly improved
  for the trace closer to clients
• High number of first-timers diminishing the hit
  ratio
• Improving caching performance
• Reorganization of the hierarchy of caches (cache
  placement)
• Caching system able to deal with the first-timers

19
Conclusions and Future Work

• Summary of main findings
• Strong variation of hit ratio along the time
• High number of first-timers (higher close to
  servers)
• Main cause of low hit ratio
• LRU policy is not able to filter the entire
  locality of a stream
• Small correlation with hit ratio
• The maximum improvement we could obtain over LRU
• less than 5 percent closer to clients
• In average 25 percent closer to servers
• Results suggest reorganization of cache topology
  and a caching system able to deal with the higher
  number of first-timers
• Future work
• Cache placement find the optimal cache
  organization in order to improve the overall
  system performance
• Auto-adaptive cache system able to minimize
  periods of poor performance

20
Questions?

• Fabricio Benevenuto, Fernando Duarte,
• Virgilio Almeida, Jussara Almeida
• fabricio, fernando, virgilio, jussara_at_dcc.ufmg.b
  r