Lessons from GiantScale Services

1
Lessons from Giant-ScaleServices

• David J. Manura
• Lehigh University, Dept. Computer Science and
  Engineering
• 2002-09-24
• (A paper presentation E. Brewer, Lessons from
  Giant Scale Services, IEEE Internet Computing,
  pp. 45-55, July/August 2001)

2
Large scale services

• ExamplesAOL Web Cache (1000 nodes 10B
  queries/day), Yahoo, Hotmail, Geocites, CNN
• ProblemsLoad management, high availability,
  evolution, growth

3
Outline

• Background
• Define a model
• Load management
• High availability
• Evolution and growth
• Summary

4
Dr. Eric Brewer

• Inktomi (www.inktomi.com)
• Products searching, caching
• Customers8 of top 10 Fortune 5009 of top 10
  U.S. universities7 of top 10 pharmaceutical
  companies

5
Simple Client-Server Model
client
client
client
IP Network
client
(www.cse.lehigh.edu?)
www.cse.lehigh.edu gt
lt 128.180.14.7
DNS
128.180.14.7
Lehigh CSE Web Server
Data
6
Basic Model
7
Model Variant 1 of 2
8
Model Variant 2 of 2
Client
Client
Client
Client
9
Traditional Switch in Network
Server
Client
(7) Application (6) Presentation (5) Session (4)
Transport (3) Network (2) Data link (1) Physical
(7) Application (6) Presentation (5) Session (4)
Transport (3) Network (2) Data link (1) Physical
Layer-3 Switch
(3) Network (2) Data link (1) Physical
10
Model Features

• Local availability
• High availability not fault tolerant
• Evolves and grows quickly
• Partial results are ok (ACID not apply)

11
Design decisions

• Load balancing method
• Replication vs. Partitioning
• Symmetry vs. Asymmetry

12
How to measure availability?

• uptime time working / total timeuptime (MTBF
  MTTR) / MTTR
• yield queries completed / queries offered
• harvest data available / complete data

13
DQ Principle

• Holding hardware, nodes, and faults constant,
• data/query queries/sec DQ constant
• Corollaryharvest yield constant

14
Graceful Degradation

• Quality reduction
• Admission control (AC)
• Reduced data freshness

15
Online Evolution

• Fast reboot
• Rolling upgrade
• Big flip

16
Lessons Learned

• ACID does not applyGraceful degredation
• Reason with the DQ Principle
• Designing for load management, availability, and
  evolution planning

17
Backup Slides
18
Coupled-Cluster
Load management
19
(No Transcript)