博文
Scalability and the Cloud
||
Scalability
- Ability to handle increasing amount of work gracefully
- Dimensions:
- Horizontal: Scaling Out
- Tradeoff between the two: Price difference; Scale up - Virtualization
YouTube
- One Billion Views per day: More than 11,000 views per second
- Video Serving
Wikipedia
- Content Serving and editing
- Reads outweigh writes
- between 25,000 and 60,000 page requests per second
Twitter, Facebook News Feed
- Messaging
- Broadcasting of messages
Components
- Load balancers: routing; could be in multi-layer; hard/soft ware
- Reverse proxy
Site: highscalability.com
https://blog.sciencenet.cn/blog-425672-322753.html
上一篇：Above the Clouds: A Berkeley View of Cloud Computing（二）
下一篇：Debate between SQL and cloud database
- Ability to handle increasing amount of work gracefully
- Dimensions:
- Load
- Vertical: Scaling Up- Handling increasing load by adding resources
- Geographic- Maintain performance in case of geographically distributed systems
- Functional- Adding new features using minimum effort
- Horizontal: Scaling Out
- Tradeoff between the two: Price difference; Scale up - Virtualization
YouTube
- One Billion Views per day: More than 11,000 views per second
- Video Serving
- Upload once, view thousands of times
- Streaming of videos
- Offline processing- Streaming of videos
Wikipedia
- Content Serving and editing
- Reads outweigh writes
- between 25,000 and 60,000 page requests per second
Twitter, Facebook News Feed
- Messaging
- Broadcasting of messages
- Propagating messages to friends
- Lots of asynchronous processing
- Lots of asynchronous processing
Components
- Load balancers: routing; could be in multi-layer; hard/soft ware
- Reverse proxy
- Software load balancing
- Caching static data: Images, Java script, Read only pages
- Operate out of memory
- Squid, HAProxy, Varnish
- Web server- Caching static data: Images, Java script, Read only pages
- Operate out of memory
- Squid, HAProxy, Varnish
- Session state
- Cookies
- Blocking connections to other tiers
- Application Server- Cookies
- Blocking connections to other tiers
- "shared nothing architecture"
- Deals with “business logic”
- Get data from DB, do some operation, send to user
- Get request from user, do some operation, store into DB
- Messaging Software- Deals with “business logic”
- Get data from DB, do some operation, send to user
- Get request from user, do some operation, store into DB
- Becoming popular among web 2.0 sites for queuing offline tasks
- Asynchronous
- Publish-Subscribe
- eg: JMS, RabbitMQ, Amazon SQS
- Caching Software- Asynchronous
- Publish-Subscribe
- eg: JMS, RabbitMQ, Amazon SQS
- What is to be cached?
- CDNs- News sites : Generated web pages
- Session state
- Static objects
- Database rows
- Any information that is expensive to read from disk, and may be accessed frequently
- eg: Memcached- Session state
- Static objects
- Database rows
- Any information that is expensive to read from disk, and may be accessed frequently
- Provide crucial usage statistics
- Optimize resource per customer
- Motivation for using CDNs
- Database- Optimize resource per customer
- Motivation for using CDNs
- Minimize bandwidth cost
- Improve latency
- Improve availability
- Akamai, Amazon Cloud Front- Improve latency
- Improve availability
- NoSQL？
- Problem: Scaling RDBMS is hard
- Solution: Don’t use RDBMS, use K-V stores
- Examples: HBase, Amazon Dynamo, BigTable, Cassandra, VoldeMort, MongoDB, CouchDB …
- Solution: Don’t use RDBMS, use K-V stores
- Examples: HBase, Amazon Dynamo, BigTable, Cassandra, VoldeMort, MongoDB, CouchDB …
Site: highscalability.com
https://blog.sciencenet.cn/blog-425672-322753.html
上一篇：Above the Clouds: A Berkeley View of Cloud Computing（二）
下一篇：Debate between SQL and cloud database
