Adaptive Configuration of a Web Caching Hierarchy Pranav A. Desai Jaspal Subhlok

1
Adaptive Configuration of a Web Caching
HierarchyPranav A. Desai Jaspal Subhlok

• Presented by
• Pranav A. Desai

2
Introduction

• Web caching
• Used to improve the performance of the World Wide
  Web
• Hierarchy of caches
• Further enhances the performance
• Goal of the research
• Improve the performance of a caching hierarchy

3
Outline

• Web caching hierarchy
• Motivation
• Approach
• Adaptive Hierarchy Management System
• Performance evaluation
• Conclusion Future work

4
Web Caching Hierarchy

• A network of cooperating caches hierarchically
  arranged in a tree-like structure
• Caches can have sibling-sibling or parent-child
  relationship with other caches

5
Web Caching Hierarchy
Parent Web Cache
parent-child relationship
Child Web Caches
sibling-sibling relationship
6
Motivation

• Limitations of a caching hierarchy
• Requires manual configuration
• Changes in network conditions may deteriorate the
  performance of the caches in the hierarchy

7
Example
Cache A
Congested network
Cache B
Cache C
All sibling hierarchy
8
Example
Cache A
Cache B
Cache C
All sibling hierarchy
9
Metrics we need to consider

• Available bandwidth (network metric)
• Indicative of the overhead associated with
  cooperating with peer caches
• Inter-cache hit ratio (cache metric)
• Measures the benefit due to hierarchy
• Other metrics that we considered
• Request hit ratio
• Request rate
• CPU load
• Service time (hits and misses)
• Round trip time

10
Solution

• Requires two components
• A mechanism
• Collect the metrics
• Reconfigure the caches
• A policy
• An algorithm that can design a hierarchy using
  the metrics

11
Adaptive Hierarchy Management System
CONTROLLER
12
Adaptive Hierarchy Algorithm

• The algorithm uses threshold values for the
  metrics to design the hierarchy
• The threshold values are determined empirically

13
Experimental Setup

• Experiments are performed on a Squid cache
  hierarchy of three sibling caches
• Bandwidth is controlled using Dummynet
• Client robots send requests from web traces
  obtained from NLANR (National Laboratory for
  Applied Network Research)
• Traces are randomly selected from different sites
  in the NLANR hierarchy

14
Determination of threshold values

• Traces used are 10000 requests long
• Bandwidth is varied in step of 100, 10, 1, 0.1
  Mbps
• To simulate realistic conditions the caches are
  warmed before performing the experiments
• Sending specific amount of requests to the caches
  before performing the experiments
• Three levels of warming 0, 50, 100
• Threshold values are determined by comparing the
  performance of the hierarchies

Hierarchy 0
Hierarchy 1
A
A
C
B
C
B
15
Impact of Sibling Cache
Hierarchy 0
Hierarchy 1
A
A
C
B
C
B

• Benefit of hierarchy is not obtained due to high
  ICP overhead and low inter-cache hit ratio

16
Impact of Sibling Cache
Hierarchy 0
A
C
B
Hierarchy 1
A
C
B
17
Adaptive Hierarchy Algorithm

• For bandwidths gt 10Mbps cooperating with peer
  caches is beneficial
• For bandwidths in the range 10 1 Mbps
  communicating with peer cache is beneficial if
  inter-cache hit ratio gt 6
• For bandwidth lt 1Mbps eliminating the
  relationship is beneficial in all cases

18
Adaptive Hierarchy Algorithm
Select a Link
BW lt 1 of maxBW ?
Set Link Relation to NONE
Set Relation to SIBLING
Y
Y
N
NONE
PARENT or SIBLING
Y
N
1 lt BW lt 10 of maxBW ?
IC_HR lt 6 ?
Check Link Relation
N
19
Performance of Adaptive Hierarchy

• Bandwidth is varied randomly in steps of 100, 10,
  1 and 0.1 Mbps
• The period for each bandwidth phase is controlled
• Each trace is about half million requests long

Sibling
A
C
B
Sibling
Sibling
20
Performance with Adaptive Hierarchy
Mean
Median

• Performance improvement of 13 and 29 is
  obtained in mean response time for cache A and
  cache B respectively
• Improvement is not evident from the median
  response time

21
Response time of individual requests
22
Conclusion

• Adaptive Hierarchy Management System is capable
  of dynamically configuring a set of caches into
  good hierarchies
• In our experimental setup Adaptive hierarchy
  performs better by around 30

23
Future Work

• Extensive evaluation of the system
• Evaluation of other metrics
• Request hit ratio
• Request rate
• Service time (Hit and Miss)
• Round trip time
• Auto discovery of caches

24
Thank You!