- Ghana • Understanding MapReduce • Map Reduce - An Introduction • Word count – default • Word count – custom Programming model to process large datasets Supported languages for MR Java Ruby Python C++ Map Reduce Programs are Inherently parallel. More data more machines to analyze. No need to change anything in the code. Start with WORDCOUNT example “Do as I say, not as I do” Word Count As 2 Do 2 I 2 Not 2 Say 1 define wordCount as Map<String,long>; for each document in documentSet { T = tokenize(document); for each token in T { wordCount[token]++; } } display(wordCount); This works until the no.of documents to process is not very large Spam filter Millions of emails Word count for analysis Working from a single computer is time consuming Rewrite the program to count form multiple machines How do we attain parallel computing ? 1. All the machines compute fraction of documents 2. Combine the results from all the machines STAGE 1 define wordCount as Map<String,long>; for each document in documentSUBSet { T = tokenize(document); for each token in T { wordCount[token]++; } } STAGE 2 define totalWordCount as Multiset; for each wordCount received from firstPhase { multisetAdd (totalWordCount, wordCount); } Display(totalWordcount) Master Documents Comp-1 Comp-2 Comp-3 Comp-4 Problems STAGE 1 • Documents segregations to be well defined Master Documents Comp-1 Comp-2 Comp-3 Comp-4 • Bottle neck in network transfer • Data-intensive processing • Not computational intensive • So better store files over processing machines • BIGGEST FLAW • Storing the words and count in memory • Disk based hash-table implementation needed Problems STAGE 2 Master • Phase 2 has only once machine • Bottle Neck • Phase 1 highly distributed though • Make phase 2 also distributed • Need changes in Phase 1 • Partition the phase-1 output (say based on first character of the word) • We have 26 machines in phase 2 • Single Disk based hash-table should be now 26 Disk based hash-table • Word count-a , worcount-b,wordcount-c Documents Comp-1 Comp-2 Comp-3 Comp-4 Master Documents Comp-1 Comp-2 Comp-3 Comp-4 A B C D E 1 2 4 5 10 Comp-10 Comp-20 A B C D E 10 20 40 5 9 . . . Comp-30 Comp-40 After phase-1 From comp-1 ▪ ▪ ▪ ▪ ▪ WordCount-A comp-10 WordCount-B comp-20 . . . Each machine in phase 1 will shuffle its output to different machines in phase 2 This is getting complicated Store files where are they are being processed Write disk-based hash table obviating RAM limitations Partition the phase-1 output Shuffle the phase-1 output and send it to appropriate reducer This is more than a lot for word count We haven’t even touched the fault tolerance What if comp-1 or com-10 fails So, A need of frame work to take care of all these things We concentrate only on business Interim output MAPPER REDUCER Comp-2 Comp-3 Comp-4 Partitioning Documents HDFS Comp-1 A B C D E 1 2 4 5 10 A B C D E 1 2 4 5 10 . . . Shuffling Master Comp-10 Comp-20 Comp-30 Comp-40 Mapper Reducer Mapper filters and transforms the input Reducer collects that and aggregate on that. Extensive research is done two arrive at two phase strategy Mapper,Reducer,Partitioner,Shuffling Work together common structure for data processing Input Output Mapper <K1,V1> List<K2,V2> Reducer <k2,list(v2)> List<k3,v3> Mapper <key,words_per_line> : Input <word,1> : output Input Output List<K2,V2> Reducer Mapper <K1,V1> <word,list(1)> : Input Reducer <k2,list(v2)> List<k3,v3> <word,count(list(1))> : Output As said, don’t store the data in memory So keys and values regularly have to be written to disk. They must be serialized. Hadoop provides its way of deserialization Any class to be key or value have to implement WRITABLE class. Java Type Hadoop Serialized Types String Text Integer IntWritable Long LongWritable Let’s try to execute the following command ▪ hadoop jar hadoop-examples-0.20.2-cdh3u4.jar wordcount ▪ hadoop jar hadoop-examples-0.20.2-cdh3u4.jar wordcount <input> <output> What does this code do ? Switch to eclipse