Evaluating Content Management Techniques for Web Proxy Caches

1
Evaluating Content Management Techniques for Web
Proxy Caches
Martin Arlitt, Ludmila Cherkasova, John Diley,
Rich Friedrich and Tai Jin (Hewlett-Packard
Laboratories) (in 2nd Workshop on Internet Server
Performance, in conjunction with ACM SIGMETRICS
99)
Cho Joon-ho(CA Lab, CS department, KAIST) 2001 .
11. 6
2
Agenda

• Problems
• Quick Tour (Summary)
• Critique
• Design Design Rationale
• Data Collection and Reduction
• Key Workload Characteristics
• Experimental Design
• Simulation Results
• Virtual Cache

3
Problems

• Current Web Proxy caches utilize simple
  replacement policies
• Relatively low hit rates
• Additional delays
• So what?
• Developing a quantitative understanding of Web
  traffic
• How effective are current proxy cache replacement
  policies for real workloads?
• Focus on two performance metrics
• Hit rate
• Byte hit rate
• Designing new replacement policies
• Utilize frequency for higher performance
• Are neither susceptible to cache pollution nor
  require parameterization

4
Agenda

• Problems
• Quick Tour (Summary)
• Critique
• Design Design Rationale
• Data Collection and Reduction
• Key Workload Characteristics
• Experimental Design
• Simulation Results
• Virtual Cache

5
Quick Tour (Summary) 1/3

• The problems of existing studies
• Short-term traces of busy proxies or long-term
  traces of relatively inactive proxies
• Long-term traces in busy environments are needed
• Trace driven simulation
• Collect total 117,652,652 requests during five
  month
• Use smaller and more compact log
• The points to be considered
• Object size
• Recency of Reference
• Frequency of Reference
• Turnover

6
Quick Tour (Summary) 2/3

• Existing replacement policy
• LRU (Least-Recently-Used)
• Size replaces the largest object
• GD-Size (GreedyDual-Size)
• Replaces the object with the lowest utility
• LFU - replaces the least frequently used object
• New replacement policy
• GDSF (GreedyDual-Size with Frequency)
• GD-Size a frequency factor
• LFU-DA (Least Frequently Used with Dynamic Aging)
• LFU-Aging a dynamic mechanism(Running age L)
• Virtual Caches
• Logically partitions the cache into N virtual
  caches

KiCi/SiL
KiFiCi/SiL
KiCiFiL
7
Quick Tour (Summary) 3/3
Analysis of Virtual Cache Performance VC0 using
GDSF-Hits, VC1 using LFU-DA
8
Agenda

• Problems
• Quick Tour (Summary)
• Critique
• Design Design Rationale
• Data Collection and Reduction
• Key Workload Characteristics
• Experimental Design
• Simulation Results
• Virtual Cache

9
Critique

• Pros
• Quantitative understanding of Web traffic
• Long term trace-driven simulation in busy proxy
  servers
• Providing two new replacement algorithms that run
  efficiently
• Providing a new cache management method, Virtual
  Cache
• Cons
• Not fresh
• No consideration of dynamic data
• No consideration of processing overhead for these
  more complex algorithms
• Performance improvements are insignificant

10
Agenda

• Problems
• Quick Tour (Summary)
• Critique
• Design Design Rationale
• Data Collection and Reduction
• Key Workload Characteristics
• Experimental Design
• Simulation Results
• Virtual Cache

11
Data Collection and Reduction

• Data collection
• Long term trace-driven simulation
• Total 117,652,652 requests were handled during
  five month period
• Data include
• Client IP address, request time, response status,
  the time required for the proxy to complete its
  response
• Data reduction
• Smaller, more compact log
• Due to storage constraint
• To ensure that analyses and simulations could be
  completed in a reasonable amount of time
• Reduction by
• Storing data in more efficient manner
• Removing information of little value

12
Key Workload Characteristics

• Cacheable Objects
• Most client requests be for cacheable objects
  (96)
• Object Set Size
• total 389GB
• Object Sizes
• Variable medium 4KB, maximum 148MB video
  clip
• Recency of reference
• 1/3 of all re-references occurred within one hour
• Frequency of reference
• Web referencing patterns are non-uniform
• Turnover
• Objects that were once popular are no longer
  requested

13
Experimental Design 1/2

• Least-Recently-Used(LRU)
• Replaces the object requested least recently
• Considers only a single work load characteristic
• Size
• Replaces the largest object
• Tries to minimize the miss ratio (target to byte
  hit rate)
• Cache pollution
• GreedyDual-Size(GD-Size)
• GD-Size(1) for Hit Rate
• GD-Size(Packets) for Byte Hit Rate

Ci the cost associated with bringing object i
into the cache Si the object size L a running
age factor
KiCi/SiL
14
Experimental Design 2/2

• LFU
• Replaces the least frequently used object
• LFU-Aging LFU Aging ? avoids cache pollution
• Parameterization problem still remains
• Greedy Dual-Size with Frequency(GDSF)
• GD-Size doesnt take into account frequency
• Least Frequently Used with Dynamic Aging(LFU-DA)
• LFU-Aging requires parameterization to perform
  well
• LFD-DA uses inflation factor as well as the
  frequency count

KiFiCi/SiL
Fi a frequency count
KiCiFiL
L a running age factor
15
Agenda

• Problems
• Quick Tour (Summary)
• Critique
• Design Design Rationale
• Data Collection and Reduction
• Key Workload Characteristics
• Experimental Design
• Simulation Results
• Virtual Cache

16
Simulation Results 1/2
17
Simulation Results 2/2
18
Agenda

• Problems
• Quick Tour (Summary)
• Critique
• Design Design Rationale
• Data Collection and Reduction
• Key Workload Characteristics
• Experimental Design
• Simulation Results
• Virtual Cache

19
Virtual Cache 1/2

• An approach that can focus on both of hit rate
  and byte hit rate simultaneously
• Mechanism
• Logically partitions the cache into N virtual
  caches
• Each virtual cache(VC)is managed with its own
  replacement policy
• Steps
• Initially all objects are in VC0
• Replacements from VCi are moved to VCi1
• Replacements from VCi1 are evicted form the
  cache
• When reaccessed, objects are reinserted in VC0

20
Virtual Cache 2/2
Figure 3. Analysis of Virtual Cache Performance
VC0 using GDSF-Hits, VC1 using LFU-DA
Figure 4. Analysis of Virtual Cache Performance
VC0 using LFU-DA, VC1 using GDSF-Hits