Scale and Performance of Cooperative Web Proxy Caching

1
Scale and Performance of Co-operative Web Proxy
Caching
Alec Wolman, Geoffrey M. Voelker, Nitin Sharma,
Neal Cardwell, Anna Karlin and Henry M. Levy
Presented by Amit Kumar Saha
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
2
What is co-operative caching ?
Network Transfer time ltlt disk access time
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
3
Co-operative Web Proxy Caching
Server
Proxy
Proxy
The INTERNET
Inter proxy communication time ltlt server fetch
time ?
Would this scale to the internet ?
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
4
Document Sharing andProxy Caching
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
5
Objectives of the simulation

• Best achievable performance
• Assignment of clients to caches
• Range of client population

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
6
Assumptions made

• Cache has infinite size
• Entries do not expire
• Include cold start misses
• Practical and ideal caches modeled

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
7
Best possible scenario
A Single Top Level Proxy
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
8
Simulation Methodology
UW trace
Simulated UW proxy
Requests
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
9
Best Achievable request rate

• Knee at 2500 clients

• Easily handled by one proxy

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
10
Scaling with population size

• Slope_Ideal(UW) Slope_Ideal(MS)

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
11
Scaling with population size

• Slope_Ideal(UW) Slope_Cacheable(UW)

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
12
Scaling with population size

• Slope_Ideal(MS) gt Slope_Cacheable(MS)

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
13
Best Achievable latency

• Constant mean

• Constant median

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
14
Best Achievable byte hit rate

• Shared objects are smaller in size

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
15
Best Achievable BW utilization

• Slope of a particular curve does not change

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
16
How much can Co-operation help ?
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
17
Simulation Methodology
UW trace
Classifier
Requests
Simulated UW proxy
Requests
Assumption Very few internal proxies
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
18
Gains from co-operative caching

• There is something to be gained before the knee

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
19
Grouped vs Random population

• Marginal increase in hit rate

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
20
Observations

• Random distribution of clients does not matter

• Grouping of clients to proxies does not help

• No results on latency

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
21
Co-operation between organizational proxies
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
22
Simulation Methodology
UW trace
Simulated UW proxy
Requests
Misses
Simulated Co-operating Proxy Preloaded with
objects of MS trace
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
23
Simulation Result

• Unpopular documents are universally unpopular

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
24
Observations

• Co-operative caching is effective for small
  population
• Little gain from organizational co-operation

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
25
An Analytical model of Web Access
Objective To analyze steady state performance
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
26
Model of Web Access

• Clients
• Number of clients
• Average client request rate
• Time distribution of requests

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
27
Model of Web Access

• Documents
• Number of documents
• Number of popular documents
• Average document size
• Average last byte latency
• Probability that a document is cacheable
• Time distribution of change of popular and
  unpopular documents

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
28
Asymptotic cacheable hit rate

• Sensitive to rate of change of documents

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
29
Asymptotic mean request latency

• Reaches mean latency of un-cacheable docs

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
30
Sensitivity to change interval

• Rate of change of unpopular documents dominate

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
31
Observations

• Small populations achieve most of the benefits of
  co-operative caching
• Average client request rate is increasing much
  faster than rate of change of unpopular documents

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
32
Implementing co-operative caching

• Hierarchical System
• Hash-based System
• Directory based System

• Queuing delays have been ignored

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
33
Hierarchical Caching System
Server
Level 1
latency
K - levels
d-1
Total population N
Each server serves (N/d level-1) sized disjoint
subsets
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
34
Hash based caching system
latency
Total population N
Servers
m d k caches
Hash function
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
35
Directory based caching system
N/m
latency
latency
N/m
Total population N
Servers
N/m
m caches
Periodic multicasts required
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
36
Latency of these schemes

• Best possible (1-pc)E(L) 0.76s

(Houston!)
Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01
37
Conclusions

• Co-operative caching benefits small population !!
• Document cacheability is crucial
• Clustering clients does not help

Cooperative Web Proxy caching, Amit Saha
Comp620, 04/04/01