Squirrel: A decentralized peer-to-peer web cache

1
Squirrel A decentralized peer-to-peer web cache

• Paul Burstein
• 10/27/2003

2
Outline

• Overview
• Design
• Evaluation
• Discussion

3
Traditional Web Caching

• Goals
• Reduce browser latency
• Reduce aggregate bandwidth
• Reduce load on web servers
• Deployment
• Dedicated centralized machines
• Placed at local network boundaries

4
Squirrel Web Caching

• Decentralized caching
• Desktops cooperate in a peer-to-peer fashion
• Mutual sharing between hosts
• Hosts browse and cache

5
Pros

• Centralized
• Dedicated Hardware
• Cost
• Administration
• Handling load bursts
• Single point of failure

• Decentralized
• No additional hardware
• More users ? more resources
• Automatic scaling
• Self organizing
• Easy deployment

6
Assumptions

• Cooperative hosts
• No security issues
• Link and node failures
• Nodes are in single geographic location
• Low internal network latencies

7
Outline

• Overview
• Design
• Evaluation
• Discussion

8
Design Goals

• Target environment 100 - 100,000 machines
• Goal Achieve performance comparable to
  centralized cache

9
Design Overview

• Built on top of Pastry
• Objects have 128-bit objectIds
• SHA-1 hash of URL
• Mapped to home node with closest nodeId
• Requests
• GET new request
• cGET conditional
• Two schemes
• Home-store
• Directory

10
Home-store

• Objects stored at client cache and home node

• External requests come through home node
• Cache replacement
• All objects are considered
• home node fresh
• home node stale

11
Directory

• Home node keeps a directory of pointers
• Randomly redirect to delegates

• no directory, add new delegate
• cGET not modified
• delegate fresh, get from delegate
• cGET and stale, update
• GET and stale, update

12
Outline

• Overview
• Design
• Evaluation
• Discussion

13
Evaluation Characteristics

• Compare two schemes and dedicated cache
• Performance
• Latency
• External bandwidth
• Hit ratio
• Overhead
• Load
• Storage
• Fault Tolerance

14
Trace Characteristics
15
Bandwidth and Hit ratio

• Bytes transferred to origin servers and back
• correlated with hit rate
• Centralized cache with infinite storage
• 100MB cache per node achieves optimal rates
• 10MB in-memory cache is reasonable
• Directory scheme
• Active nodes suffer from eviction
• Distributed LRU is worse than centralized
• Home-store
• More total storage required

16
Latency

• User-perceived time for a response
• With comparable hit ratios, only consider
  internal hops
• Many requests can be satisfied locally, with 0
  hops
• Directory scheme latency is up to one hop greater
• Some requests can be satisfied by home node
• Squirrel Latency
• Based on Pastry hops on cache hit
• Overshadowed on cache miss

17
Load on Nodes(1/2)

• Bursty behavior observations
• Max objects served per second
• Up to 48 and 55 objects per second served for the
  two traces
• Directory scheme
• One delegate can get bombarded with requests from
  many home nodes
• Home-store scheme
• Replicate objects at request threshold

18
Load on Nodes(2/2)

• Sustained load measurements
• Max objects/minute
• Average load in any second or minute
• 0.31 objects/minute
• Redmond trace, both models

19
Fault Tolerance

• Internet connection loss
• Internal partitioning
• Individual failure
• Desktop shutdown or reboot
• Graceful shutdown
• Pastry aided content transfer
• Directory scheme
• More vulnerable to failures

20
Results

• The home-store models seems to outperform the
  directory model
• Hit ratio
• Load balancing
• Internal network latency
• Compared to centralized cache?

21
Outline

• Overview
• Design
• Evaluation
• Discussion

22
Discussion

• Would this be deployed in a corporate network?