1. Leroy Garcia

2. What is Map Reduce?  A patented programming model developed by Google Derived from LISP and other forms of functional programming  Used for processing large data and generating large data sets  Exploits large set of commodity computers  Executes process in distributed manner  Easy to use, no messy code

3. Implementation at Google  Machines w/ Multiple Processors  Commodity Networking Hardware  Cluster of Hundreds or Thousands of Machines  IDE Disks used for storage  Input Data managed by GFS  Users submit jobs to a scheduling system

4. Introduction  How does Map Reduce work?

5. Overview  Programming Model  Implementation  Refinement  Performance  Related Topics  Conclusion

6. Programming Model

7.  Map  Input: key/value pair  Key: ex. Document Name  Value: ex. Document Contents  Output:  Set of Intermediate key/values

8. Programming Model  Reduce  Input: Intermediate key, values  Key: ex. A Word  Values: Values  Output  List of Values or a Single Value

9. REDUCEREDUCE MAP Partitioning Function

10. Execution MMMMMMM k1:v k1:v k2:vk1:vk3:v k2:vk4:v k5:vk4:vk1:v k3:v Group by Key k1:v,v,v,vk2:vk3:v,vk5:v RRRR Input: Intermediate: Grouped: k4:v,v,v R

11. Parallel Execution MMMMMMM k1:v k1:v k2:vk1:vk3:v k2:vk4:v k5:vk4:vk1:v k3:v k1:v,v,v,v k2:v k3:v,v k5:v R R R RR k4:v,v,v Map Task 1Map Task 2Map Task 3 Sort and Group Partition Function Reduce 1

12. The Map Step v k kv kv map v k v k … kv Input key-value pairs Intermediate key-value pairs … kv

13. Reduce Step kv … kv kv kv Intermediate key-value pairs group reduce kvkvkv … kv … kv kvv vv Key-value groups Output key-value pairs

14. Word Count MAP v {Girl,3} {Girl,8} {Girl,5} {Boy,34} {Boy,16} {Boy,12} {Boy,23} {Girl,18} {Boy,34} {Boy,12} {Boy,23} {Boy,16} {Girl,18} {Girl,5} {Girl,8} {Girl,5} {Girl,12} {Boy,85} {Girl,43}

15. Examples  Distributed Grep  Count of URL Access Frequency  Reverse Web-Link Graph  Term-Vector per Host  Inverted Index  Distributed Sort

16. Practical Examples  Large PDF Generation  Artificial Intelligence  Statistical Data  Geographical Data

17. Large-Scale PDF Generation The New York Times needs to generate PDF files for 11,000,000 articles (every article from 1851-1980) in the form of images scanned from the original paper Each article is composed of numerous TIFF images which are scaled and glued together Code for generating a PDF is relatively straightforward

18.  Compute statistics Central Limit Theorem  N voting nodes cast votes (map)  Tally votes and take action (reduce) Artificial Intelligence

19. Statistical Analysis Photos from: stockcharts.com Statistical analysis of current stock against historical data Each node (map) computes similarity and ROI. Tally Votes (reduce) to generate expected ROI and standard deviation

20. Geographical Data  Large data sets including road, intersection, and feature data  Problems that Google Maps has used MapReduce to solve Locating roads connected to a given intersection Rendering of map tiles Finding nearest feature to a given address or location

21.  Input: Graph describing node network with all gas stations marked  Map: Search five mile radius of each gas station and mark distance to each node  Sort: Sort by key  Reduce: For each node, emit path and gas station with the shortest distance  Output: Graph marked and nearest gas station to each node Geographical Data

22. Implementation

23. Map/Reduce Walkthrough  Map: (Functional Programming)uses a function on each element of the array  Mapper: The node that performs a function on one element of the set.  Reduce: (Functional programming) iterate a function across an array  Reducer: The node that reduces across all the like-keyed elements.

24. Execution Overview 1. Split input files 2. Starts up copies of the program in cluster. Copy of program is sent to the Master Master assigns either map or reduce responsibilities 3. Map Worker reads the splits Parses key/value pairs out of the input data Passes each pair to the user-defined Map function. 4. Buffer pairs are written to local disc partitioned into regions by partitioning function The locations of these buffered pairs on the local disk are passed back to the master. Master is responsible for forwarding these locations to the reduce workers. 5. Location of the buffer pairs are given to Reduce Worker by the master Sorts Intermediate keys 6. The reduce worker iterates over the sorted intermediate data for each unique intermediate key. passes the key and the corresponding set of intermediate values to the user's Reduce function. The output of the Reduce function is appended to a final output file for this reduce partition. 7. When all map tasks and reduce tasks have been completed, the master wakes up the user program

25. Distributed Execution Overview User Program Worker Master Worker fork assign map assign reduce read local write remote read, sort Output File 0 Output File 1 write Split 0 Split 1 Split 2 Input Data

26. Fault Tolerance  Worker Failure  Master Failure  Dealing with Stragglers  Locality  Task Granularity  Skipping Bad Record

27. Worker Failure Master Worker A Map Task 1 Complete Worker B Map Task 2 Complete Failed Map Task 2 Idle Worker C Map Task 2 In Progress Worker AZ Reduce Task 1 Failed Ping Reduce Task 1 Idle Worker BX Reduce Task 1 In Progress Ping

28. Master Failure Checkpoints Checkpoint 124 Checkpoint 123 Checkpoint 125 Master Fail MASTER Checkpoint 125 NEW MASTER

29. Dealing with Stragglers  Straggler- a machine in a cluster than is running significantly slower than the rest Straggler Map Task Good Machine Map Task Copy Finish Task Line

30. Locality  Input Data is stored locally  GFS divides files in 64 MB blocks  Stores 3 copies of the blocks on different machines  Finds Replica of input data and scheduled map tasks.  Map tasks scheduled so GFS input block replica are on same machine or same rack

31. Task Granularity  Minimizes time for fault recovery  Can pipeline shuffling with map execution  Better dynamic load balancing  Often use 200,000 map/5000 reduce tasks w/ 2000 machines

32. Refinements

33. Partitoning Function The users of MapReduce specify the number of reduce tasks/output files that they desire. Data gets partitioned across these tasks using a partitioning function on the intermediate key. Special partitioning function. eg.hash(Hostname(urlkey)) mod R. Ordering Guarantee Intermediate keys are process in increasing key order. Generates sorted output per partition.

34. Combiner Function(Optional) Used by the Map Task when there is a significant repetition in the intermediate keys produced by each Map Task Map Worker Map Function Combiner Function Text Document (Girls, 1) (Girls, 2) (Girls, 6)

35. Input and Output Types Input: Supports reading data of various formats Support for new input type using a simple implementation of a reader interface. Ex.Database Ex. Datastructure Mapped in Memory Output: User codes supports to handle new type

36. Skipping Bad Records Map/Reduce functions sometimes fail for particular inputs Best solution is to debug & fix, but not always possible On seg 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

37. Status Pages

40. Performance Tests run on cluster of 1800 machines: 4 GB of memory Dual-processor 2 GHz Xeons with Hyperthreading Dual 160 GB IDE disks Gigabit Ethernet per machine Bisection bandwidth approximately 100 Gbps MR_Grep Scan 10 10 100-byte records to extract records matching a rare pattern (92K matching records) MR_Sort Sort 10 10 100-byte records (modeled after TeraSort benchmark) Two Benchmarks

41. MR_Grep Locality optimization helps: 1800 machines read 1 TB of data at peak of ~31 GB/s Without this, rack switches would limit to 10 GB/s Inputs Scanned

42. MR_Sort NormalNo Backup Tasks200 Processes Killed

43. Related Topics

44. Hadoop – Open-source implementation of MapReduce  HDFS Primary storage system used by Hadoop applications. HDFS creates multiple replicas of data blocks and distributes them on compute nodes throughout a cluster to enable reliable, extremely rapid computations. Amazon Elastic Compute Cloud (EC2) – Virtualized computing environment designed for use with other Amazon services (especially S3) Amazon Simple Storage Service (S3) – Scalable, inexpensive internet storage which can store and retrieve any amount of data at any time from anywhere on the web – Asynchronous, decentralized system which aims to reduce scaling bottlenecks and single points of failure Other Notable Implementations of MapReduce

45. Conclusion MapReduce has proven to be a useful abstraction Greatly simplifies large-scale computations at Google Easily Handles machine failure. Allows users to focus on problem, without having to deal with complicated code behind the scene.

46. Questions?????