Distributed Data Processing on the Cloud

Size: px

Start display at page:

Download "Distributed Data Processing on the Cloud"

Ashlee Patterson
5 years ago
Views:

1 Distributed Data Processing on the Cloud LTAT Lecture 1: Data analytics in the cloud Satish Srirama

2 Outline Introduction to the course Cloud computing overview Big data & analytics Distributed data processing 9/7/2018 Satish Srirama 2/50

3 Course Purpose Introduce cloud computing concepts Introduce data analytics in the cloud Introduce Distributed computing algorithms such as MapReduce BigDatasolutions such as Pig and Spark NoSQL Glance of research at Mobile & Cloud Lab in cloud computing domain 9/7/2018 Satish Srirama 3/50

4 Course Schedule Lectures: (Weeks 1-16) Fri (J. Liivi 2-122) Practice sessions: (Weeks 2-16) Mon (Ülikooli17-115) Tue (Ülikooli17-115) 9/7/2018 Satish Srirama 4/50

5 Related Courses Computer Science MSc Curriculum Distributed Systems module - Specialization course MTAT Large-scale Data Processing on the Cloud Until Fall 2017, being discontinued from this year MTAT Basics of Cloud Computing (3 ECTS) Spring 2019 MTAT Mobile and Cloud Computing Seminar (3 ECTS) Wed , Ülikooli MTAT Mobile Application Development Projects (3 ECTS) Thu , Ülikooli /7/2018 Satish Srirama 5/50

6 Questions Is everyone comfortable with data structures? How comfortable you are with algorithms? How comfortable you are with programming? Java? External APIs? Web programming Python We will also briefly touch SQL, Rand JavaScriptin labs 9/7/2018 Satish Srirama 6/50

7 Grading Written exam 50% Labs 45% ~14 lab exercises You must submit lab exercise solutions by Monday 23:55 of the next week Late submission for a week (75%) Late submission for 2 weeks (50%) Hard deadlines will be mentioned later Active participation in the lectures (Max 5%) To pass the course You need to score at least 50% in each of the subsections You need to score at least 50% in the total 9/7/2018 Satish Srirama 7/50

8 Course schedule 07.09Data analytics in the cloud Introduction to MapReduce MapReduce algorithms Information Retrieval with MapReduce Graph Data Processing with MapReduce Joins with MapReduce and MapReduce limitations Higher level scripting languages for distributed data processing In-memory data processing SQL abstraction for distributed data processing 09.11DataFrameabstraction for distributed data processing Graph processing with Bulk Synchronous Parallel model Stream Data Processing Structured Streaming Non-Relational databases Distributed Machine Learning and R Examination 1 9/7/2018 Satish Srirama 8/50

9 Course schedule -continued Labs Follow lectures icals 9/7/2018 Satish Srirama 9/50

Reference Books Mastering Cloud Computing: Foundations and

Authors: Jimmy Lin and Chris Dyer http://lintool.github.

10 Reference Books Mastering Cloud Computing: Foundations and Applications Programming Authors: Rajkumar Buyya, Christian Vecchiola, S.Thamarai Selvi Data-Intensive Text Processing with MapReduce Authors: Jimmy Lin and Chris Dyer /MapReduce-book-final.pdf White, Tom. Hadoop: the definitive guide. O'Reilly, /7/2018 Satish Srirama 10/50

11 Reference Books -continued Holden Karau, Andy Konwinski, Patrick Wendell, MateiZaharia(2015.), Learning Spark, "O'Reilly Media, Inc." 9/7/2018 Satish Srirama 11/50

12 Reference Papers M. Armbrustet al., Above the Clouds, A Berkeley View of Cloud Computing, Technical Report, University of California, Feb, Dean, J., & Ghemawat, S. (2008). MapReduce: simplified data processing on large clusters. Communications of the ACM, 51(1), Relevant papers will be mentioned in respective lectures 9/7/2018 Satish Srirama 12/50

13 CLOUD COMPUTING OVERVIEW 9/7/2018 Satish Srirama 13

It s nothing new It s a trap...we ve redefined Cloud It s worse than stupidity: it s Computing to include everything marketing hype. Somebody is that we already do.

.. other than change likely to be a set of businesses the wording of some of our ads. campaigning to make it true. WHAT IS CLOUD COMPUTING?

14 It s nothing new It s a trap...we ve redefined Cloud It s worse than stupidity: it s Computing to include everything marketing hype. Somebody is that we already do... I don t saying this is inevitable and understand what we would do whenever you hear that, it s very differently... other than change likely to be a set of businesses the wording of some of our ads. campaigning to make it true. WHAT IS CLOUD COMPUTING? Larry Ellison, CEO, Oracle (Wall Richard Stallman, Founder, Free Street Journal, Sept. 26, 2008) Software Foundation (The Guardian, Sept. 29, 2008) No consistent answer! Everyone thinks it is something else 9/7/2018 Satish Srirama 14 Slide taken from Professor Anthony D. Joseph s lecture at RWTH Aachen

15 What is Cloud Computing? Computing as a utility Utility services e.g. water, electricity, gas etc. Consumers pay based on their usage 9/7/2018 Satish Srirama 15/50

16 Clouds -Why Now (not then)? Experience with very large datacenters Unprecedented economies of scale Transfer of risk Technology factors Pervasive broadband Internet Maturity in Virtualization Technology Business factors Minimal capital expenditure Pay-as-you-go billing model 9/7/2018 Satish Srirama 16/50

17 Virtualization Virtualization techniques are the basis of the cloud computing Virtualization technologies partition hardware and thus provide flexible and scalable computing platforms App App App Virtual machine techniques VMware and Xen OpenNebula Amazon EC2 OS OS Hypervisor Hardware OS Virtualized Stack 9/7/2018 Satish Srirama 17/50

18 Containers Use kernel of the host operating system instead of a hypervisor Linux namespace Lightweight No hypervisor overhead Each Container acquires only required resources Fast start-up/ Bootup Starting a container is faster than booting new OS or spinning up a new VM Performance Near native performance Different container frameworks LXC, Docker, Linux VServer, OpenVZ Docker is an Open platform [ 9/7/2018 Satish Srirama 18/50

19 Cloud Computing -Characteristics Illusion of infinite resources No up-front cost Fine-grained billing (e.g. hourly) Gartner: Cloud computing is a style of computing where massively scalable IT-related capabilities are provided as a service across the Internet to multiple external customers 9/7/2018 Satish Srirama 19/50

20 Cloud Computing -Services Software as a Service SaaS A way to access applications hosted on the web through your web browser Platform as a Service PaaS Provides a computing platform and a solution stack (e.g. LAMP) as a service Infrastructure as a Service IaaS Use of commodity computers, distributed across Internet, to perform parallel processing, distributed storage, indexing and mining of data Virtualization SaaS Facebook, Flikr,Myspace.com, Google maps API, Gmail PaaS Google App Engine, Force.com, Hadoop, Azure, Heroku, etc IaaS Amazon EC2, Rackspace, GoGrid, SciCloud, etc. Level of Abstraction 9/7/2018 Satish Srirama 20/50

21 Cloud Computing -Themes Massively scalable On-demand & dynamic Only use what you need -Elastic No upfront commitments, use on short term basis Accessible via Internet, location independent Transparent Complexity concealed from users, virtualized, abstracted Service oriented Easy to use SLAs SLA Service Level Agreement 9/7/2018 Satish Srirama 21/50

Cloud Models Internal (private) cloud Cloud with in an organization Community cloud Cloud infrastructure jointly owned by several organizations Public cloud

22 Cloud Models Internal (private) cloud Cloud with in an organization Community cloud Cloud infrastructure jointly owned by several organizations Public cloud Cloud infrastructure owned by an organization, provided to general public as service Hybrid cloud Composition of two or more cloud models 9/7/2018 Satish Srirama 22/50

23 Short Term Implications of Clouds Startups and prototyping Minimize infrastructure risk Lower cost of entry Batch jobs One-off tasks Washington post, NY Times Cost associatively for scientific applications Research at scale 9/7/2018 Satish Srirama 23/50

24 Economics of Cloud Users Many cloud applications have cyclical demand curves Daily, weekly, monthly, Pay by use instead of provisioning for peak Resources Capacity Demand Resources Capacity Demand Time Time Static data center Data center in the cloud Unused resources 9/7/2018 Satish Srirama 24/50

25 Economics of Cloud Users -continued Risk of over-provisioning: underutilization Huge sunk cost in infrastructure Capacity Unused resources Resources Demand Time Static data center 9/7/2018 Satish Srirama 25/50

26 Economics of Cloud Users -continued Heavy penalty for under-provisioning Resources Time (days) Capacity Demand Resources Resources Time (days) Lost revenue Time (days) Lost users 9/7/2018 Satish Srirama 26/50 Capacity Demand Capacity Demand

27 Long Term Implications of clouds Application software: Cloud & client parts, disconnection tolerance Infrastructure software: Resource accounting, VM awareness Hardware systems: Containers, energy proportionality 9/7/2018 Satish Srirama 27/50

28 Cloud Computing Progress [Armando Fox, 2010] 9/7/2018 Satish Srirama 28/50

29 BIG DATA & ANALYTICS 9/7/2018 Satish Srirama 29

30 Our Data-driven World Already as of 2012, every day 2.5 exabytes( ) of data were created [ ] By 2025, IDC predicts there will be 163 zettabytes (10 21 ) of data Science Data bases from astronomy, genomics, environmental data, Sensors Humanities and Social Sciences Scanned books, historical documents, social interactions data, Business & Commerce Corporate sales, stock market transactions, census, airline traffic, Entertainment Internet images, MP3 files, Medicine MRI & CT scans, patient records, Internet of Things 50 billion devices will be connected by year 2020 [Cisco] 9/7/2018 Satish Srirama 30/50

Scientific breakthroughs Business process

31 What can we do with this wealth? What can we do? Scientific breakthroughs Business process efficiencies Realistic special effects Improve quality-of-life: healthcare, transportation, environmental disasters, daily life, Could We Do More? YES: but need major advances in our capability to analyze this data [Agrawal et al, EDBT 2011 Tutorial] 9/7/2018 Satish Srirama 31

32 Centralized Data Processing Traditionally all the data is moved to centralized storage for processing E.g. Relational database management systems (RDBMS) Data warehousing solutions What are the advantages? Economies of scale (equipment and personnel) Lack of duplication Ease in enforcing standards, security What are the disadvantages? Can the big data be stored in a single data base or for that matter in a single cluster or datacenter? 9/7/2018 Satish Srirama 32/50

33 Distributed Data Processing Computers are dispersed throughout organization Can they be used for data processing? What are the advantages? Allows greater flexibility in meeting individual needs More redundancy Responsiveness Availability Resource Sharing Increased User Involvement & Control 9/7/2018 Satish Srirama 33/50

34 Distributed Data Processing - drawbacks More difficult to test & failure diagnosis More components and dependence on communication means more points of failure Incompatibility of components Incompatibility of data More complex management & control Duplication of effort 9/7/2018 Satish Srirama 34/50

35 Platforms for Data Analysis Data Warehousing, Data Analytics & Decision Support Systems Data Analytics in the Web Context Data Analytics in the Cloud 9/7/2018 Satish Srirama 35/50

36 Data Warehousing, Data Analytics & Decision Support Systems Used to manage and control business Transactional Data: historical or point-in-time Optimized for inquiry rather than update Use of the system is loosely defined and can be ad-hoc Used by managers and analysts to understand the business and make judgments Limited scalability [Agrawal et al, EDBT 2011 Tutorial] 9/7/2018 Satish Srirama 36/50

37 Information Retrieval & Data Analytics in the Web Context (Lecture 4) Source Selection Resource Query Formulation Query Search Results System discovery Vocabulary discovery Concept discovery Document discovery Selection Documents Examination Information source reselection Delivery 9/7/2018 Satish Srirama 37

38 Data Analytics in the Cloud Scalability to large data volumes: Scan 100 TB on 1 50 MB/sec = 23 days Scan on 1000-node cluster = 33 minutes Divide-And-Conquer (i.e., data partitioning) This is where MapReduce jumps in Replicate data and computation (Lectures 2-6) Cost-efficiency: Commodity nodes (cheap, but unreliable) Commodity network Automatic fault-tolerance 9/7/2018 Satish Srirama 38/50

39 BigDataSolutions We mainly study Hadoop MapReduce Hadoop MapReduce is one of the most used solutions for large scale data analytics Spark will be discussed as an alternative (Lectures 8 and above) In-Memory MapReduce However, Hadoop has some troubles Writing low level MapReducecode is slow Need a lot of expertise to optimize MapReducecode Prototyping requires compilation A lot of custom code is required Hard to manage more complex MapReduce job chains 9/7/2018 Satish Srirama 39/50

40 Pig (Lecture 7) Pig is a high level language on top of Hadoop MapReduce Models a scripting language (Latin) Fast prototyping Similar to declarative SQL Easier to get started In comparison to Hadoop MapReduce: 5% of the code 5% of the time 9/7/2018 Satish Srirama 40/50

41 Scaling Information Systems Fault tolerance, high availability & scalability are essential prerequisites for any enterprise application deployment Scalability Generally nodes in information systems support specific load When load increases beyond certain level, systems should be scaled up Similarly when load decreases they should be scaled down 9/7/2018 Satish Srirama 41/50

42 Scaling Information Systems - continued Two basic models of scaling Vertical scaling Also known as Scale-up Achieving better performance by replacing an existing node with a much powerful machine Risk of losing currently running jobs Horizontal scaling Aka Scale-out Achieving better performance by adding more nodes to the system New servers are introduced to the system to run along with the existing servers 9/7/2018 Satish Srirama 42/50

43 Scaling Enterprise Applications in the Cloud Client Site Client Site Client Site Load Balancer (Proxy) App Server App Server App Server App Server App Server DB Memcache 9/7/2018 Satish Srirama 43

44 Load Balancer Load balancing has been a key mechanism in making efficient web server farms Load balancer automatically distributes incoming application traffic across multiple servers Some of the key load balancing algorithms include Random, Round Robin, Least connection Examples: Nginx- HAProxy- Pen - Basics of Cloud Computing - MTAT (Spring 2019) 9/7/2018 Satish Srirama 44/50

Server App Server Database becomes the DB Memcache Scalability

45 Scaling in the Cloud -bottleneck Client Site Client Site Client Site Load Balancer (Proxy) App Server App Server App Server App Server App Server Database becomes the DB Memcache Scalability Bottleneck Cannot leverage elasticity 9/7/2018 Satish Srirama 45

Memcache but limited consistency and operational

46 Scaling in the Cloud -bottleneck Client Site Client Site Client Site Load Balancer (Proxy) App Server App Server App Server App Server App Server Scalable and Elastic, Key Value DB Stores Memcache but limited consistency and operational flexibility NoSQL 9/7/2018 Satish Srirama (Lecture 14) 46

47 Other topics of interest Distributed machine learning (Lecture 15) 9/7/2018 Satish Srirama 47/50

48 This week in Lab (Homework) Introduction to IaaS & Getting keys Work with Virtual Machines 9/7/2018 Satish Srirama 48/50

49 Next lecture Introduction to MapReduce 9/7/2018 Satish Srirama 49/50

50 References Several of the slides are taken from Prof. Anthony D. Joseph s lecture at RWTH Aachen (March 2010) Some slides are also taken from Big Data and Cloud Computing: Current State and Future Opportunities, tutorial at EDBT 2011 by D. Agrawal, S. Das and A. Abbadi. Dave Heroy, Distributed data processing, M. Armbrustet al., Above the Clouds, A Berkeley View of Cloud Computing, Technical Report, University of California, Feb, Rajkumar Buyya, Satish Narayana Srirama, Giuliano Casale, et al, : A Manifesto for Future Generation Cloud Computing: Research Directions for the Next Decade, ACM Computing Surveys, ISSN , ACM Press, New York, USA. (in press, accepted on July 20, 2018). 9/7/2018 Satish Srirama 50/50

51 NoSQL If you want vast, on-demand scalability, you need a non-relational database. Since scalability requirements: Can change very quickly and, Can grow very rapidly. Difficult to manage with a single in-house RDBMS server RDBMS scale well: When limited to a single node, but Overwhelming complexity to scale on multiple server nodes Solution: NoSQL(Lecture 6) 9/7/2018 Satish Srirama 51/50

52 Algorithms for Massive Data Problems (Lecture 8) We conclude the course with some massively parallel data algorithms Streaming algorithms 9/7/2018 Satish Srirama 52/50

Mobile Cloud Computing

MTAT.03.262 -Mobile Application Development Lecture 8 Mobile Cloud Computing Satish Srirama, Huber Flores satish.srirama@ut.ee Outline Cloud Computing Mobile Cloud Access schemes HomeAssignment3 10/20/2014