谷歌工程师讲解云计算视频：http://v.youku.com/v_show/id_XMTEwMDEwOTU2.html

叶平是 Google 的软件工程师。他于 1996 年在台湾大学物理系取得博士学位，主修实验粒子物理学。于 1998 年起在中研院担任助研究员，2001 年换至台湾大学高能物理实验室任客座及兼任助理教授。自 1993 年起，他参与过美国费米加速器实验室、美国太空总署、欧洲粒子物理中心、日本高能加速器研究机构的粒子物理实验或天文粒子观测，之后于 2007 年夏天加入 Google。叶平目前是 Google 台湾繁体中文搜寻品质工作及大学计划和台湾云计算计划的负责人。

转自：http://www.programmer.com.cn/515/

＝＝＝＝＝＝＝＝

Evolution of Computing:

• network computing; -- Network is Computer.(client/server); Seperation of Funtionalities
  
• cluster computing; -- Tightly coupled compputing resources: CPU, Storage, Data (usually within a LAN)
  
• grid computing; -- Resource sharing across domains; Decentralized; Open standards
  
• utility computing. -- Don't buy computers, lease computing powering; Ownership model
  

Next step: Cloud computing

Definition: services and data, put in the cloud, accessible, scalability

Powers of Cloud Computing:

1. commodity hardware;
2. infrastucture software:

• GFS;
• BigTable;
• MapReduce

Web Application:

• Serving;
• Database;
• Storage;
• Data Processing

Google's Solution:

• Goolge AppEngine;
• BigTable;
• GFS;
• MapReduce

Implimentation:

• Hadoop -- GFS and MapReduce
• HBase -- BigTable
• HDFS -- GFS
• ...

GFS:

• Files are broken into chunks;
• Chunks triplicated across three different servers;
• Metadata managed by Master; (chunkId, path, serverId, etc)
• Data transfers happen directly between clients and chunk servers

GFS used in Google:

• 200+ clusters;
• File system clusters of up to 5000+ machine;
• Pools of 10000+ clients;
• Read is Faster than Write;

BigTable:

• 3-dimension: row, column, timestamp;
• Distributed;
  
• Scalable;
• Self-Managing -- add/remove servers; load balance;
• System structure: -- BigTable Cell (ie, cluster)
  
• BigTable client library
  
• Master; -- metadata operations; load balance
  
• Tablet Server; -- serves data
  
• Cluster scheduling system -- handle failover, monitoring;
  
• GFS -- holds tablet data, logs;
  
• Lock service -- hold metadata, master election
• Currently -500 cells

Distributed Data Processing:

• File Management: where to store files? Distributed; -- Master
  
• Granularity: Splitting;
• Job Allocation: assign which task to which node? prefer local job;
• Fault-recovery: what if a node crashes?Redundancy, Crash-detection, Job re-allocation;

• Map: (in_key, in_value) --> ( (keyj, value) | j = 1, 2, ...)
• Reduce: (key, (value1, value2, ... )) --> (key, f_value)

• 200,000 mappers;
• 500 reducers;
• 2000 nodes

Data Playground:

Summary:
Cloud Computing is about  scalable web applications and data processing.

Reference:
http://v.youku.com/v_show/id_XMTEwMDEwOTU2.html