1. By: Jeffrey Dean & Sanjay Ghemawat Presented by: Warunika Ranaweera Supervised by: Dr. Nalin Ranasinghe

2. MapReduce: Simplified Data Processing on Large Clusters In Proceedings of the 6th Symposium on Operating Systems Design and Implementation (OSDI' 04) Also appears in the Communications of the ACM (2008)

3.  Ph.D. in Computer Science – University of Washington  Google Fellow in Systems and Infrastructure Group  ACM Fellow  Research Areas: Distributed Systems and Parallel Computing

4.  Ph.D. in Computer Science – Massachusetts Institute of Technology  Google Fellow  Research Areas: Distributed Systems and Parallel Computing

5.  Calculate 30*50 Easy?  30*50 + 31*51 + 32*52 + 33*52 +.... + 40*60 Little bit hard?

6.  Simple computation, but huge data set  Real world example for large computations  20+ billion web pages * 20kB webpage  One computer reads 30/35 MB/sec from disc  Nearly four months to read the web

7.  Parallelize tasks in a distributed computing environment  Web page problem solved in 3 hours with 1000 machines

8. o How to parallelize the computation? o Coordinate with other nodes o Handling failures o Preserve bandwidth o Load balancing  Complexities in Distributed Computing

9.  A platform to hide the messy details of distributed computing  Which are,  Parallelization  Fault-tolerance  Data distribution  Load Balancing  A programming model  An implementation

10.  Example: Word count the quick brown fox the fox ate the mouse the 1 quick 1 brown 1 fox 1 the 1 fox 1 ate 1 the 1 mouse 1 the 3 quick 1 brown 1 fox 2 ate 1 mouse 1 DocumentMappedReduced the 1

11.  Eg: Word count using MapReduce the quick brown fox the fox ate the mouse Map Reduce the, 1 quick, 1 brown, 1 fox, 1 the, 1 fox, 1 ate,1 the, 1 mouse, 1 Input Map Reduce Output the, 3 quick, 1 brown, 1 fox, 2 ate, 1 mouse, 1

12. map(String key, String value): for each word w in value: EmitIntermediate(w, "1"); Intermediate key/value pair – Eg: (“fox”, “1”) Document Name Document Contents Input  Text file Output  (“fox”, “1”)

13. reduce(String key, Iterator values): int result = 0; for each v in values: result += ParseInt(v); Emit(AsString(result)); Word List of Counts (Output from Map) Input  (“fox”, {“1”, “1”}) Output  (“fox”, “2”) Accumulated Count

14.  Reverse Web-Link Graph Target (My web page) Target (My web page) Source Web page 2 Source Web page 2 Source Web page 3 Source Web page 3 Source Web page 4 Source Web page 4 Source Web page 5 Source Web page 5 Source Web page 1 Source Web page 1

15.  Reverse Web-Link Graph Map (“My Web”, “Source 1”) (“Not My Web”, “Source 2”) (“My Web”, “Source 3”) (“My Web”, “Source 4”) (“My Web”, “Source 5”) Reduce (“My Web”, {“Source 1”, “Source 3”,.....}) Target Source pointing to the target Source web pages

16. Implementation: Execution Overview Map LayerReduce Layer User Program Master Worker Input Layer Intermediate Files Output Layer Split 1 Split 2 Split 3 Split 4 Split 0 (1) Fork (2) Assign Map (2) Assign Reduce (3) Read(4) Local Write (5) Remote Read O/P File 1 O/P File 0 (6) Write

17.  Complexities in Distributed Computing, to be solved o How to parallelize the computation? o Coordinate with other nodes o Handling failures o Preserve bandwidth o Load balancing o Automatic parallelization using Map & Reduce o How to parallelize the computation?

18.  Restricted Programming model  User specified Map & Reduce functions  1000s of workers, different data sets Worker1 Worker2 Worker3 User-defined Map/Reduce Instruction Data

19. o Automatic parallelization using Map & Reduce  Complexities in Distributed Computing, solving.. o Coordinate with other nodes o Handling failures o Preserve bandwidth o Load balancing o Coordinate nodes using a master node

20.  Master data structure  Pushing information (meta-data) between workers Information Master Map Worker Reduce Worker

21. o Fault tolerance (Re-execution) & back up tasks o Coordinate nodes using a master node o Automatic parallelization using Map & Reduce  Complexities in Distributed Computing, solving.. o Handling failures o Preserve bandwidth o Load balancing

22.  No response from a worker task?  If an ongoing Map or Reduce task: Re-execute  If a completed Map task: Re-execute  If a completed Reduce task: Remain untouched  Master failure (unlikely)  Restart

23.  “Straggler”: machine that takes a long time to complete the last steps in the computation  Solution: Redundant Execution  Near end of phase, spawn backup copies  Task that finishes first "wins"

24. o Saves bandwidth through locality o Fault tolerance (Re-execution) & back up tasks o Coordinate nodes using a master node o Automatic parallelization using Map & Reduce  Complexities in Distributed Computing, solving.. o Preserve bandwidth o Load balancing

25.  Same data set in different machines  If a task has data locally, no need to access other nodes

26. o Saves bandwidth through locality o Fault tolerance & back up tasks o Coordinate nodes using a master node o Automatic parallelization using Map & Reduce  Complexities in Distributed Computing, solving.. o Load balancing o Load balancing through granularity  Complexities in Distributed Computing, solved

27.  Fine granularity tasks: map tasks > machines  1 worker  several tasks  Idle workers are quickly assigned to work

28.  Partitioning  Combining  Skipping bad records  Debuggers – local execution  Counters

29. Normal Execution No backup tasks 891 S1283 S  44% increment in time  Very long tail  Stragglers take >300s to finish

30. Normal Execution 200 processes killed 891 S 933 S  5% increment in time  Quick failure recovery

31.  Clustering for Google News and Google Product Search  Google Maps  Locating addresses  Map tiles rendering  Google PageRank  Localized Search

32.  Apache Hadoop MapReduce  Hadoop Distributed File System (HDFS)  Used in,  Yahoo! Search  Facebook  Amazon  Twitter  Google

33.  Higher level languages/systems based on Hadoop  Amazon Elastic MapReduce  Available for general public  Process data in the cloud  Pig and Hive

34.  Large variety of problems can be expressed as Map & Reduce  Restricted programming model  Easy to hide details of distributed computing  Achieved scalability & programming efficiency

35.  GFS solution: Shadow masters  Only meta data is passed through the master  A new copy can be started from the last point of state

36.  Programmer’s burden?  “If we hadn’t had to deal with failures, if we had a perfectly reliable set of computers to run this on, we would probably never have implemented MapReduce” – Sanjay Ghemawat

37. Combiner