Healing the Web: An Overview of CoDeeN

1
Healing the WebAn Overview of CoDeeN Related
Projects

• Vivek Pai, Larry Peterson many others
• Princeton University

2
How This All Started

• OSDI paper on Content Distribution Network (CDN)
  algorithms
• NS-2 simulations on 8GB AlphaServer
• Limits of detailed simulation 70 hrs per
  experiment, huge memory requirements
• Still only 1 as large as real networks
• PlanetLab developed
• Suddenly, can get real traffic

3
Toward a Self-Healing Web

• Goals
• Improve all levels of Web infrastructure
• Interoperate where possible
• Drive with real traffic
• Feed back measurements

Client
Replication
Data Transfer
Name Lookup
Routing
4
Basic CoDeeN
Data Transfer
origin
CoDeeN Proxy
Each CoDeeN proxy is a forward proxy, reverse
proxy, redirector
5
By The Numbers
Data Transfer

• Running 24/7 since June 2003
• Over 1 million unique IPs as clients
• About 20 billion requests serviced
• Over 25 million reqs/day
• Highest-request project on PlanetLab
• 2 in bytes/day we serve smaller objs
• not including PlanetLab Dec 2003 upgrade

6
CoDNS
Name Lookup
My LAN
Client Programs
CoDNS
My Machine
7
Availability
Name Lookup

• Adds one 9, from 99 to 99.9

CoDNS
LDNS
8
Evaluation Live Traffic
Name Lookup

9
CoBlitz
Replication
CDN reverse caches the chunks!
chunk1
chunk2
CDN
CDN
CDN Redirector Reverse Proxy
chunk3
Agent
CDN
Client
CDN
Client
Agent
CDN
CDN
chunk5
chunk4
10
CoBlitz Performance
Replication

• CoBlitz users
• CiteSeer digital library
• Fedora Core Linux
• PlanetLab projects

11
PlanetSeer
Routing

• Host service
• CDN
• Components
• Passive monitoring
• Active probing
• Advantages
• Low overhead
• Wide coverage

Client
C
R1
R2
B
A
12
Confirmed Anomaly Breakdown
Routing
Temp Anomalies 16

• Confirmed anomalies
• 271,898 in 90 days
• 2 per minute
• 100x more

Persist Loop 7
Temp loop 1
Path Change 44
Other Outage 23
Fwd Outage 9
13
Going Forward

• Cycles bandwidth latency
• qualitatively change opportunities
• Obvious gains performance, capacity, storage
• Byzantine-like protocols become feasible
• More time for reasoning online failure
  detection, repair, recovery
• Deeper analysis better classifiers, boosting
• Better security more cycles needed for crypto,
  dynamic SSL