1. C-Store: MapReduce Jianlin Feng School of Software SUN YAT-SEN UNIVERSITY May. 22, 2009

2. Motivation: Large Scale Data Processing Many tasks:  Process lots of data to produce other data. The Problems  How to parallelize the computation?  How to distribute the data?  How to handle failures? MapReduce is a programming model for solving the above problems.

3. Programming model Input & Output: each a set of key/value pairs Programmer specifies two functions:  map (in_key, in_value) -> list(out_key, intermediate_value) Processes input key/value pair Produces set of intermediate pairs  reduce (out_key, list(intermediate_value)) -> list(out_value) Combines all intermediate values for a particular key Produces a set of merged output values (usually just one)

4. Example: Count word occurrences map(String input_key, String input_value): // input_key: document name // input_value: document contents for each word w in input_value: EmitIntermediate(w, "1"); reduce(String output_key, Iterator intermediate_values): // output_key: a word // output_values: a list of counts int result = 0; for each v in intermediate_values: result += ParseInt(v); Emit(AsString(result));

5. Implementation Environment at Google Large clusters of commodity PCs connected together with switched Ethernet.  100s/1000s of 2-CPU x86 machines, 2-4 GB of memory  Commodity networking hardware: 100 MB/s or 1GB/s.  Storage is on local IDE disks  GFS: distributed file system manages data (SOSP'03)  Job scheduling system: jobs made up of tasks, scheduler assigns tasks to machines Implementation is a C++ library linked into user programs

6. Typical Hadoop Cluster (Hadoop is developed mainly by Yahoo!)

7. 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

8. Data flow Input, final output are stored on a distributed file system  Scheduler tries to schedule map tasks “ close ” to physical storage location of input data Intermediate results are stored on local FS of map and reduce workers Output is often input to another map reduce task

9. Coordination Master data structures  Task status: (idle, in-progress, completed)  Idle tasks get scheduled as workers become available  When a map task completes, it sends the master the location and sizes of its R intermediate files, one for each reducer  Master pushes this info to reducers Master pings workers periodically to detect failures

10. Failures Map worker failure  Map tasks completed or in-progress at worker are reset to idle  Reduce workers are notified when task is rescheduled on another worker Reduce worker failure  Only in-progress tasks are reset to idle Master failure  MapReduce task is aborted and client is notified

11. How many Map and Reduce jobs? M map tasks, R reduce tasks Rule of thumb:  Make M and R much larger than the number of nodes in cluster  One DFS chunk per map is common  Improves dynamic load balancing and speeds recovery from worker failure Usually R is smaller than M, because output is spread across R files

12. Combiners Often a map task will produce many pairs of the form (k,v1), (k,v2), … for the same key k  E.g., popular words in Word Count Can save network time by pre-aggregating at mapper  combine(k1, list(v1))  v2  Usually same as reduce function Works only if reduce function is commutative and associative

13. Partition Function Inputs to map tasks are created by contiguous splits of input file For reduce, we need to ensure that records with the same intermediate key end up at the same worker System uses a default partition function e.g., hash(key) mod R Sometimes useful to override  E.g., hash(hostname(URL)) mod R ensures URLs from a host end up in the same output file

14. Implementations Google  Not available outside Google Hadoop  An open-source implementation in Java  Uses HDFS for stable storage  Download: http://lucene.apache.org/hadoop/ http://lucene.apache.org/hadoop/

15. References Jeffrey Dean and Sanjay Ghemawat, MapReduce: Simplified Data Processing on Large Clusters. http://labs.google.com/papers/mapreduce.html The HTML Slides. http://labs.google.com/papers/mapreduce-osdi04- slides/index.html http://labs.google.com/papers/mapreduce-osdi04- slides/index.html Matei Zaharia. Introduction to MapReduce and Hadoop. www.cs.berkeley.edu/~demmel/cs267_Spr09/Lectures/C loud_MapReduce_Zaharia.ppt The Stanford CS345A Slides on MapReduce. http://www.stanford.edu/class/cs345a/mapreduce.ppt