Web Prefetching: Costs, Benefits and Performance

1
Web Prefetching Costs, Benefits and Performance

• Yingyin Jiang, Min-You Wu, Wei Shu
• Dept. of Electrical Computer Engineering
• The University of New Mexico
• Aug 15, 2002
• WCW 2002, Boulder, Colorado

2
Talk Focus

• A solution space of web prefetching
• Several object-selection criteria
• New simple selection algorithm
• Select a good set of objects
• Maximize benefit/cost
• Can be tuned to achieve different goals
• New performance metric
• Balance of benefits (hit rate improvement)
  against costs(network bandwidth increase)

3
Outline

• Motivation
• Our approach
• Performance Evaluation
• Discussions and Conclusions

4
Motivation for Prefetching

• Limited hit ratio by passive caching
• Typical -- 20 to 40
• Limited by newly introduced, dynamically
  generated data and rapid changes of objects in
  the web
• Prefetching can further improve hit ratio (reduce
  client latency) but sacrifice network bandwidth
• Predict future accesses to objects
• Fetch objects before users request them

5
Key Parameters for Prefetching Algorithms

• Object popularity
• Zipf popularity distribution
• pi C / i a
• The probability of a request for the ith most
  popular document is inversely proportional to i
• Object lifetime
• Indication of object modification
• Key factor for the design of prefetching
  algorithm

6
Solution Space for Web Prefetching

• Six models
• Two extreme cases
• Passive caches(non-prefetching)
• Prefetching all objects
• The other four algorithms use different
  object-selecting criteria and fetch objects with
  values that exceed the threshold
• Popularity
• Lifetime
• Good Fetch
• APL

7
Two Simple Schemes

• Popularity
• Keep the most popular objects in the system
• Update these objects immediately whenever they
  are modified
• Threshold objects popularity
• Lifetime
• Keep objects with longest lifetimes
• Mostly consider the network resource demands
• Threshold the expected lifetime of object

8
Good Fetch

• Proposed by Venkataramani in Venkataramani01
• Attempt to balance the benefit against the cost
  of keeping an object
• Threshold probability that a prefetched object
  is accessed before it changes
• 
  
• 
  ,
• li object is expected lifetime
• a avg. request arrival rate
• pi object is popularity
• Prefetch object i if

9
APL

• Attempt to balance benefit against cost
• Threshold the expected number of requests for
  the object i that arrive during its lifetime
• ,
• li object is expected lifetime
• a avg. request arrival rate
• pi object is popularity
• Prefetch object i if

10
Enhanced APL

• Enhanced APL
• Prefetch object i if
• Motivation -- adapt to network status
• When the network has abundant bandwidth, a larger
  value of n can be used to fetch more popular
  objects to improve hit ratio
• When the network has congestions, a smaller value
  of n can be used or prefetching can even be
  disabled to save the bandwidth

11
Performance Evaluation for Prefetching

• Evaluation metrics
• Benefit -- hit ratio
• Cost bandwidth
• Benefit/cost H/B
• Algorithms to be evaluated
• Popularity
• Lifetime
• Good Fetch
• APL

12
New Evaluation Metric H/B

• Measure benefit/cost
• Passive caching serves as a baseline for
  comparison
• Enhanced
• Emphasize benefit -- hit ratio improvement
• When system has plenty of spare bandwidth, a
  small fraction of hit ratio improvement can still
  be justified

13
Evaluation Methodology

• Analytical simulation evaluation
• Give a proof of concept for performance of
  different algorithms
• Experimental settings
• Poisson model of user request arrival
• Workload of one million objects Douglis97,
  Breslau99, Nishikawa98
• Zipf popularity distribution, with parameter
  0.986 Breslau99
• Object lifetimes distribution obtained from
  Douglis97
• Fixed object size 10K Bytes Bray96,
  Williams96, Abdulla98, Arlitt99
• No correlation between lifetimes, sizes,
  popularities Crovella98, Breslau99

14
Distribution of Object Lifetimes
Douglis97

• We vary the mean lifetime of objects across
  several orders of magnitude
• Shifting factor 0, mean 3.8 months
• Shifting factor -2, mean 1.2 days
• Shifting factor -4, mean 16. 7 minutes
• The shift factor denotes the horizontal
  displacement along the lifetime axis (on log
  scale) of the Cumulative Distribution Function

15
Results -- Hit Ratio
shift factor -4
shift factor 0
Hit Ratio
Hit Ratio
Log10( of prefetched objects)
Log10( of prefetched objects)

• Popularity -- the highest hit ratio
• APL (n 1) works very close to GoodFetch
• GoodFetch and APL (n 1) work closer to Lifetime
  at longer mean lifetime, and closer to Popularity
  at shorter mean lifetime
• Lifetime the lowest hit ratio

16
Results -- Bandwidth
shift factor -4
BW(kbps)
BW(kbps)
Log10( of prefetched objects)
Log10( of prefetched objects)

• Popularity consumes the most network bandwidth
  compared to others
• GoodFetch and APL obtain significant improvement
  in hit ratio at an expense of moderate bandwidth
  increase
• e.g. when prefetching 0.1 objects, 15 increase
  on hit ratio (39.54 over 24.3)
• Total bandwidth lt 2demand bandwidth (113.50 kbps
  over 60.57 kbps)
• Lifetime consumes a smallest amount of bandwidth

17
Results -- H/B
shift factor 0
shift factor -4
H/B
H/B
Log10( of prefetched objects)
Log10( of prefetched objects)

• Popularity drop quickly, not comparable with
  others
• GoodFetch and APL -- attain high H/B values and
  show their effectiveness on maximizing
  benefit/cost
• Lifetime -- slowly decrease all the way from the
  beginning

18
APL Family Hit Ratio
Hit Ratio
Log10( of prefetched objects)

• a ( pi )n li , n 0.5, 1, 2, 5
• n 1, APL -gt Good Fetch, maximize benefit/cost
• n gt 1, APL -gt Popularity, increase hit ratio
• n lt 1, APL -gt Lifetime, reduce bandwidth
  consumption

19
APL Family Bandwidth
BW(kbps)
Log10( of prefetched objects)

• E.g., n 5, APLs hit ratio is very close to
  Popularity, and bandwidth cost is favorably
  smaller.

20
APL Family H/B Ratio
H/B
Log10( of prefetched objects)

• From H/B point of view, APL can bias on popular
  or long-lived objects without sacrificing too
  much benefit/cost
• n 1, achieve the best benefit/cost
• n gt 1, get more increase on hit ratio with fair
  BW consumption
• n lt 1, reduce bandwidth and still with reasonable
  hit ratio

21
Enhanced Hk/B

• Recall -- why do we extend H/B to Hk/B?
• Emphasize hit ratio improvement when evaluating
  benefit/cost
• When evaluating with Hk/B, a small fraction of
  hit ratio improvement can still be justified even
  at the cost of disproportionate bandwidth
  increase

22
Hk/B Evaluation
Hk/B
Log10( of prefetched objects)

• With higher k, it allows prefetching of more
  objects -gt encourage more hit ratio improvement
  with Hk/B
• For Hk/B gt 1, how many objects can be prefetched?
• K 1 --700 k 2 -- 2,000 k 3 -- 7,000 k
  4 -- 30,000 k 5 -- 200,000

23
Discussions
100
Hit Ratio
Bandwidth

• We obtain a solution space for prefetching where
  different strategies lie along axes of hit ratio
  and bandwidth with different performance

24
Conclusions

• We propose a new prefetching algorithm APnL
  that can be made adaptive to different network
  status by varying n
• Prefetching must consider both object popularity
  and lifetime in order to significantly improve
  hit ratios at modest costs