1. Introduction to Cloud Computing 彭波
北京大学信息科学技术学院
5/25/2009
First, ask for opinions on the future of computing technology, in 5-10 years
Mobile computing, grid computing, Artificial intelligence, cloud computing or something…
What the computer looks like at that time
How do you use your computer at that time
These are interesting and challenging questions.

2. 大纲 What is Cloud Computing? Build a big cloud Why we are here?
What we will get from this class?
Background and our target.

3. 云计算(Cloud Computing) 教学方法：故事，video，pictures  得到（隐藏的）概念，留下印象
Tell the story from
SaaS
PaaS
Utility Computing
 Cloud Computing
Story about
Key player’s story
User experiences
Slogan: Cloud Computing =
2 为什么要Cloud Computing?
Some key characteristics
3.Cloud Computing面临哪些问题？
讨论，imagination for your own dream
人们的评论: positive and negative , objective 认识

4. What is Cloud Computing?
First write down your own opinion about “cloud computing” , whatever you thought about in your mind.
Question: What ? Who? Why? How? Pros and cons?
The most important question is: What is the relation with me?
Story line:高清《云计算》最浅显解谜云故事.flv
First write down your own opinion about “cloud computing” 写下来whatever you though about in your mind question: what ? Who? Why? How? Pros and cons? The most important question is : the relation with me ?
Watch the video take the question and search the answer
Let’s find out the key concept in the story!
SaaS
PaaS
Utility Computing
 Cloud Computing

5. Cloud Computing is… No software access everywhere by Internet
power -- Large-scale data processing
Appeal for startups
Cost efficiency
实在是太方便了
Software as platform
Cons
Security
Data lock-in
SaaS
PaaS
Utility Computing

6. Software as a Service (SaaS)
a model of software deployment whereby a provider licenses an application to customers for use as a service on demand.
.vs. 传统的EULA(End User License Agreement )
Level4, Scalable, Configurable, Multi-Tenant-Efficient：增加multitier architecutre，支持load-balanced farm of identical application instances，在数量变化的多个服务器上运行。Provider按用户需求demand来增减服务器，而不需要修改任何软件的体系结构。

7. Platform as a Service (PaaS)
对于开发Web Application和Services，PaaS提供了一整套基于Internet的，从开发，测试，部署，运营到维护的全方位的集成环境。特别它从一开始就具备了Multi-tenant architecture，用户不需要考虑多用户并发的问题，而由platform来解决，包括并发管理，扩展性，失效恢复，安全。
LAMP is the industry standard
But management is a hassle:
Configuration, tuning
Backup and recovery, disk space management
Hardware failures, system crashes
Software updates, security patches
Log rotation, cron jobs, and much more
Redesign needed once your database exceeds one box

8. Utility Computing
“pay-as-you-go” 好比让用户把电源插头插在墙上，你得到的电压和Microsoft得到的一样，只是你用得少，pay less；utility computing的目标就是让计算资源也具有这样的服务能力，用户可以使用500强公司所拥有的计算资源，只是use less pay less。这是cloud computing的一个重要方面
好比让用户把电源插头插在墙上，你得到的电压和Microsoft得到的一样，只是你用得少，pay less；utility computing的目标就是让计算资源也具有这样的服务能力，用户可以使用500强公司所拥有的计算资源，只是use less pay less。这是cloud computing的一个重要方面

9. Cloud Computing is…

10. Key Characteristics
illusion of infinite computing resources available on demand;
elimination of an up-front commitment by Cloud users; 创业启动花费
ability to pay for use of computing resources on a short-term basis as needed。小时间片的billing，报告指出utility computing在这一点上的实践是失败的
Cloud Computing Provider需要具备的能力包括：very large datacenters, large-scale software infrastructure以及operational expertise来运行它们。这对于Google, Amazon, Microsoft, eBay这样的公司来说是必然的事。另一个重要的驱动力来自于very large datacenter比medium-sized datacenter有5-7倍的经济效益。认为构建extremely large-scale的commodity-computer datacenters是cloud computing的一个关键enabler。
very large datacenters
large-scale software infrastructure
operational expertise

11. Why now? very large-scale datacenter的实践， 因为新的技术趋势和Business模式
pay-as-you-go computing
伴随Web2.0的出现，过去”high-touch, high-margin, high-commitment”的服务提供开始转变为”low-touch, low-margin, low-commitment”的self-service。比如credit card在以前需要与payment processing service(VeriSign)签订合同，这对于个人和小business来说很困难。当PayPal出现时，任何个人可以用credit card来支付，不需要contract，没有long-term commitment，”touch”（客户支持和关系管理）的开销几乎没有

12. Key Players Amazon Web Services Google App Engine
Microsoft Windows Azure

13. Key Applications
Mobile Interactive applications, Tim O’Reilly相信未来是属于能够实时对用户提供信息的服务。Mobile必定是关键。而后台在datacenter中运行是很自然的模式，特别是那些mashup融合类型的服务。
Parallel batch processing。大规模数据处理使用Cloud Computing技术很自然，MapReduce，Hadoop在这里起到重要作用。这里，数据移入/移出cloud是很大的开销，Amazon开始尝试host large public datasets for free。
The rise of analytics。数据库应用中transaction based应用还在增长，而analytics的应用增长迅速。数据挖掘，用户行为分析等应用的巨大推动。
Extension of compute-intensive desktop application。计算密集型的任务，说matlab, mathematica都有了cloud computing的扩展，woo~

14. Cloud Computing = Silver Bullet?
Google文档在3月7日发生了大批用户文件外泄事件。美国隐私保护组织就此提请政府对Google采取措施，使其加强云计算产品的安全性。
Problem of Data Lock-in

15. Challenges
Refer [1]

16. Some other Voices
The interesting thing about Cloud Computing is that we’ve redefined Cloud Computing to include everything that we already do I don’t understand what we would do differently in the light of Cloud Computing other than change the wording of some of our ads.
Larry Ellison, quoted in the Wall Street Journal, September 26, 2008
It’s stupidity. It’s worse than stupidity: it’s a marketing hype campaign. Somebody is saying this is inevitable — and whenever you hear somebody saying that, it’s very likely to be a set of businesses campaigning to make it true.
Richard Stallman, quoted in The Guardian, September 29, 2008

17. What’s matter with ME?!
What you want to do with 1000pcs, or even 100,000 pcs?
What you want to do with 100,000 pcs?
It’s really possible now!
Calculate Amaon EC2 service price for this:

18. Cloud is coming…

19. Build a big “Cloud”
In distributed system course, we learn much to deal with how
To build large scale distributed system.
Can we use them to build a big cloud to became a cloud provider?

20. First let’s look inside the requirement of some emerging applications: large-scale data processing
slides borrowed from
Data-Rich Computing: Where It’s At @hadoop summit 2008
Phillip B. GibbonsIntel Research Pittsburgh

21. Example: Wikipedia Anthropology
Kittur, Suh, Pendleton (UCLA, PARC), “He Says, She Says: Conflict and Coordination in Wikipedia” CHI, 2007
Increasing fraction of edits are for
work indirectly related to articles
Experiment
Download entire revision history of Wikipedia
4.7 M pages, 58 M revisions, 800 GB
Analyze editing patterns & trends
Computation
Hadoop on 20-machine cluster
3 slides from
“Data Intensive Super Computing”
Randal E. Bryant, Carnegie Mellon University

22. Example: Scene Completion
Hays, Efros (CMU), “Scene Completion Using Millions of Photographs” SIGGRAPH, 2007
Image Database Grouped by Semantic Content
30 different Flickr.com groups
2.3 M images total (396 GB).
Select Candidate Images Most Suitable for Filling Hole
Classify images with gist scene detector [Torralba]
Color similarity
Local context matching
Computation
Index images offline
50 min. scene matching, 20 min. local matching, 4 min. compositing
Reduces to 5 minutes total by using 5 machines
Extension
Flickr.com has over 500 million images …

23. Example: Web Page Analysis
Fetterly, Manasse, Najork, Wiener (Microsoft, HP), “A Large-Scale Study of the Evolution of Web Pages,” Software-Practice & Experience, 2004
Experiment
Use web crawler to gather 151M HTML pages weekly 11 times
Generated 1.2 TB log information
Analyze page statistics and change frequencies
Systems Challenge
“Moreover, we experienced a catastrophic disk failure during the third crawl, causing us to lose a quarter of the logs of that crawl.”

24. Let’s build a big Computer…
Given datacenter with tens of thousands of pcs, can you make all these tasks easier and run faster?
Software infrastructure 的关键部件是？
Distributed storage system
Storage: file system or database system
Programming Framework: parallel programming
Challenges:
scalability
fault-tolerance
Distributed Computing Framework

25. Challenges 大规模数据处理面临的困难 大规模PC机群scaling reliably is hard!
On 1000s of nodes
MTBF < 1 day
With so many disks, nodes, switches something is always broken
并行/分布式程序开发，调试is hard!
数据如何划分
任务如何调度
任务之间的通信
错误处理，容错…
Storage System &
Computing Framework
良好可扩展性
良好的容错能力
What have we learned to deal with these problems?
replications and consistency
fault-tolerance, recovery by logging & checkpoints
consensus,
Programming Model
一定的表达能力
很好的简单易用性

26. Cluster-Based Distributed File Systems
Observation: When dealing with very large data collections, following a simple client-server approach is not going to work.
Solution 1: For speeding up file accesses, apply striping techniques by which files can be fetched in parallel:
(a) whole-file distribution, (b) file-striped system

27. A natural DFS design
File stripping as Chunks

28. Master of DFS 功能 问题 元数据管理
inode:　file -> <filename, timestamp, size, owner, chunklist…>
运行数据管理
Chunk server info管理：map(chunk, chunkserver)
Client info管理: locks, open files, etc. 问题
Performance bottleneck?
Master failure?
Master Recovery?
A p2p(dHT) design to calculate chunk address to chunk id is very interesting.
Multiple master with data space partition is scalable, but not safe and even more fragile
Metadata all in memory leads to high performance, -> support less files, it that ok?
But when crash, data lost is inevitable?

29. ChunkServer of DFS 功能 问题 管理chunk data: chunkid -> local file
Performance bottleneck?
Chunkserver failure -> data lost?
Replications is one answer

30. Review on DFS design Workload Goal Strategies 大数据 顺序读和append操作为主
Reliability, availability, scalability…
Tolerance to hardware failures
Managing numerous files of large size
Optimizing commonly performed operations
Strategies
Chunk Replications (fault tolerance and performance)
Large chunk size (MB)
All metadata in memory on Master, with operation log

31. Data Replications in DFS
Master
Chunkserver
/foo/bar.dat
Chunkserver
Chunkserver
Client
Chunkserver
Chunkserver
Chunkserver

32. Data Mutations Two kinds of data mutations are supported Chunk A B
Random writes
Record appends
Leases used to maintain consistent mutation order A B
Replica A B
Record appends much more frequent than random writes
Concurrent appends are common
Replica A B

33. Primary-based Consistency Protocol
What if a mutation operation fail in the
middle?
Chunkserver
Chunkserver
Secondary replica
Master
Chunkserver
Primary replica
Chunkserver
Secondary replica
并发append时，primary replica的网络带宽成为瓶颈
Any other design? 取消并发写！
Chunkserver
Client
Chunkserver
/foo/bar.dat

34. Relaxed Consistency Model
修改操作后的文件区域状态
Consistent
不管从那个replicas读，所有clients看到相同数据
Defined
consistent + 所有clients看到更新操作写入的全部数据
Undefined
consistent +但是可能不能反映任意一个更新操作写入的数据
Inconsistent
Clients不同时间看到不同的数据
Refer to [3]
GFS employs a relaxed consistency model
GFS defines two types of mutations
Writes cause data to be written to a specific location specified by the application
Record appends cause data to be written at an offset chosen by GFS. GFS ensures that data is written atomically at least once. I will talk about the append-at-least-once semantics later.
GFS’s consistency model is based on the definition of states of a file region after mutation
A file region is said to be consistent if all clients see the same data, regardless which replicas they read from.
In addition, if all clients see what the mutation writes in its entirety, it is defined. A file region is in undefined state if it is consistent but it doesn’t reflect what any one mutation has written
Otherwise, it will be inconsistent, that is, different clients see data at different times
Relaxed consistency model
Two types of mutations
Writes
Cause data to be written at an application-specified file offset
Record appends
Operations that append data to a file
Cause data to be appended atomically at least once
Offset chosen by GFS, not by the client

35. Consistency Model (contd)
不提供完全严格的一致性[3]
由应用程序处理这种放宽的一致性下出现的inconsistent数据区域问题
提供atomic append，保证append at least once
A file region can have different states. A file region is said to be in a consistent state, if all the clients see the same data irrespective of the chunkserver they read the data from. A file region is said to be in the defined state after a mutation, if all the clients see the same data, irrespective of the chunkserver they read the data from, and if the clients can see that the mutation has written in its entirety. The table shows the state of the region for the different cases. For a data write, if it is serial (i.e. only one client is writing at any time to the region) and if the write succeeds, then the region is in a defined state. In the case of concurrent writes to the region, if the write succeeds, it is in a consistent state but it is undefined, because the region would contain fragments of different writes. In the case of record appends, the state is defined in both the serial and the concurrent cases. But the defined regions are interspersed with inconsistent regions. If the mutation operation fails, the region is inconsistent. The application is supposed to take care of the relaxed consistency. The applications do this by writing self identifying records.

36. Summary for DFS Architecture: master-worker
File strip : large chunk size
Scalability & Availability: Chunk replication
Primary-based consistency protocol
Relaxed consistency model

37. Distributed Computing
大规模机群 + 可靠存储（DFS）上怎样计算？ 编程 运行 调试

38. Example: Web Page Analysis
Fetterly, Manasse, Najork, Wiener (Microsoft, HP), “A Large-Scale Study of the Evolution of Web Pages,” Software-Practice & Experience, 2004
Experiment
Use web crawler to gather 151M HTML pages weekly 11 times
Generated 1.2 TB log information
Analyze page statistics and change frequencies
Systems Challenge
“Moreover, we experienced a catastrophic disk failure during the third crawl, causing us to lose a quarter of the logs of that crawl.”
map on chunks (data blocks)
Reduce: summary the result
Independent tasks

39. A simple solution
M：提取网页长度，按domain执行数据合并

40. A possible solution
M: 提取网页长度,按domain执行数据合并
R: 按domain执行数据合并

41. A More difficult Problem
统计文档集中每个word出现的次数?

42. Shuffle Implementation

43. Partition and Sort Group
Partition function: hash(key)%reducer number
Group function: sort by key

44. A Distributed Computing Framework
Parallel/Distributed Computing Programming Model
I’m the
MapReduce Framework
Input split
shuffle
output

45. Typical problem solved by MapReduce
读入数据: key/value 对的记录格式数据
Map: 从每个记录里extract something
map (in_key, in_value) -> list(out_key, intermediate_value)
处理input key/value pair
输出中间结果key/value pairs
Shuffle: 混排交换数据
把相同key的中间结果汇集到相同节点上
Reduce: aggregate, summarize, filter, etc.
reduce (out_key, list(intermediate_value)) -> list(out_value)
归并某一个key的所有values，进行计算
输出合并的计算结果 (usually just one)
输出结果
Outline stays the same, map and reduce change to fit the problem

46. Mapreduce Framework

47. Word Frequencies in Web pages
输入：one document per record
用户实现map function，输入为
key = document URL
value = document contents
map输出 (potentially many) key/value pairs.
对document中每一个出现的词，输出一个记录<word, “1”>

48. Example continued: MapReduce运行系统(库)把所有相同key的记录收集到一起 (shuffle/sort)
用户实现reduce function对一个key对应的values计算
求和sum
Reduce输出<key, sum>

49. Model is Widely Applicable MapReduce Programs In Google Source Tree
Refer to [4]
Example uses:
distributed grep
distributed sort
web link-graph reversal
term-vector / host
web access log stats
inverted index construction
document clustering
machine learning
statistical machine translation
...

50. Algorithms Fit in MapReduce
文献中见到实现了的算法
K-Means, EM, SVM, PCA, Linear Regression, Naïve Bayes, Logistic Regression, Neural Network
PageRank
Word Co-occurrence Matrices，Pairwise Document Similarity
Monte Carlo simulation ……

51. Capability of MapReduce
解决大部分并行计算需求的主要手段？
MapReduce难于有效实现的并行算法[2]
Dense/Sparse Linear Algebra
N-Body Problems
Dynamic Programming
Graph Traversal
Combinational Logic
。。。
Ucb/Eecs, K. Asanovic, R. Bodik, B. Catanzaro, J. Gebis, P. Husbands, K. Keutzer, D. Patterson, W. Plishker, J. Shalf, S. Williams, and K. Yelick, "The landscape of parallel computing research: a view from Berkeley,"
"The landscape of parallel computing
research: a view from Berkeley," 2006

52. Google MapReduce Architecture
Single Master node
Many worker bees
Many worker bees
How is this distributed?
Partition input key/value pairs into chunks, run map() tasks in parallel
After all map()s are complete, consolidate all emitted values for each unique emitted key
Now partition space of output map keys, and run reduce() in parallel
If map() or reduce() fails, reexecute!

53. MapReduce Operation Master informed of result locations
Initial data split
into 64MB blocks
M sends data location to R workers
Computed, results locally stored
Final output written

54. Fault Tolerance 通过re-execution实现fault tolerance
周期性heartbeats检测failure
Re-execute失效节点上已经完成+正在执行的 map tasks
Why????
Re-execute失效节点上正在执行的reduce tasks
Task completion committed through master
Robust: lost 1600/1800 machines once  finished ok
Master Failure?

55. Refinement: Redundant Execution
Slow workers significantly delay completion time
Other jobs consuming resources on machine
Bad disks w/ soft errors transfer data slowly
Solution: Near end of phase, spawn backup tasks
Whichever one finishes first "wins"
Dramatically shortens job completion time

56. Refinement: Locality Optimization
Master scheduling policy:
Asks GFS for locations of replicas of input file blocks
Map tasks typically split into 64MB (GFS block size)
Map tasks scheduled so GFS input block replica are on same machine or same rack
Effect
Thousands of machines read input at local disk speed
Without this, rack switches limit read rate

57. Refinement: Skipping Bad Records
Map/Reduce functions sometimes fail for particular inputs
Best solution is to debug & fix
Not always possible ~ third-party source libraries
On segmentation fault:
Send UDP packet to master from signal handler
Include sequence number of record being processed
If master sees two failures for same record:
Next worker is told to skip the record

58. Other Refinements Compression of intermediate data Combiner
“Combiner” functions can run on same machine as a mapper
Causes a mini-reduce phase to occur before the real reduce phase, to save bandwidth
Local execution for debugging/testing
User-defined counters

59. Summary CloudComputing brings How to make it real?
Possible of using unlimited resources on-demand, and by anytime and anywhere
Possible of construct and deploy applications automatically scale to tens of thousands computers
Possible of construct and run programs dealing with prodigious volume of data …
How to make it real?
Distributed File System
Distributed Computing Framework
…………………………………

60. Q&A

61. 参考文献
[1] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. H. Katz, A. Konwinski, G. Lee, D. A. Patterson, A. Rabkin, I. Stoica, and M. Zaharia, "Above the Clouds: A Berkeley View of Cloud Computing," EECS Department, University of California, Berkeley UCB/EECS , February
[2] Ucb/Eecs, K. Asanovic, R. Bodik, B. Catanzaro, J. Gebis, P. Husbands, K. Keutzer, D. Patterson, W. Plishker, J. Shalf, S. Williams, and K. Yelick, "The landscape of parallel computing research: a view from Berkeley,"
[3] G. Sanjay, G. Howard, and L. Shun-Tak, "The Google file system," in Proceedings of the nineteenth ACM symposium on Operating systems principles. Bolton Landing, NY, USA: ACM Press, 2003.
[4] J. D. a. S. Ghemawat, "MapReduce: Simplified Data Processing on Large Clusters," in Osdi, 2004, pp

62. Google App Engine App Engine handles HTTP(S) requests, nothing else
Think RPC: request in, processing, response out
Works well for the web and AJAX; also for other services
App configuration is dead simple
No performance tuning needed
Everything is built to scale
“infinite” number of apps, requests/sec, storage capacity
APIs are simple, stupid

63. App Engine Architecture
req/resp
stateless APIs
R/O FS
urlfech
Python VM
process
stdlib
mail
app
images
datastore
stateful
APIs
memcache 63

64. Microsoft Windows Azure

65. Amazon Web Services
Amazon’s infrastructure (auto scaling, load balancing)
Elastic Compute Cloud (EC2) – scalable virtual private server instances
Simple Storage Service (S3)
Simple Queue Service (SQS) – messaging
SimpleDB - database
Flexible Payments Service, Mechanical Turk, CloudFront, etc.

66. Amazon Web Services
Very flexible, lower-level offering (closer to hardware) = more possibilities, higher performing
Runs platform you provide (machine images)
Supports all major web languages
Industry-standard services (move off AWS easily)
Require much more work, longer time-to-market
Deployment scripts, configuring images, etc.
Various libraries and GUI plug-ins make AWS do help
Use AWS if you:
Want to use third party open source software
Have existing code
Want to transfer web app to own machine/servers later on
Port code to another language
Want control (EC2 instances and machine images)
Ex: need to stress/load test app – just load up 1000 instances
Features – messaging, payment services, etc

67. Price of Amazon EC2