1. Jeffrey D. Ullman Stanford University

2. 2 Chunking Replication Distribution on Racks

3.  Datasets can be very large.  Tens to hundreds of terabytes.  Cannot process on a single server.  Standard architecture emerging:  Cluster of commodity Linux nodes (compute nodes).  Gigabit Ethernet interconnect.  How to organize computations on this architecture?  Mask issues such as hardware failure.

4. Mem Disk CPU Mem Disk CPU … Switch Each rack contains 16-64 nodes Mem Disk CPU Mem Disk CPU … Switch 1 Gbps between any pair of nodes in a rack 2-10 Gbps backbone between racks

5.  First order problem: if nodes can fail, how can we store data persistently?  Answer: Distributed File System.  Provides global file namespace.  Examples: Google GFS, Colossus; Hadoop HDFS.  Typical usage pattern:  Huge files.  Data is rarely updated in place.  Reads and appends are common.

6.  Chunk Servers.  File is split into contiguous chunks, typically 64MB.  Each chunk replicated (usually 2x or 3x).  Try to keep replicas in different racks.  Alternative: Erasure coding.  Master Node for a file.  Stores metadata, location of all chunks.  Possibly replicated.

7. 7  Organized into racks.  Intra-rack connection typically gigabit speed.  Inter-rack connection faster by a small factor.

8. 8 Racks of Compute Nodes File Chunks

9. 9 3-way replication of files, with copies on different racks.

10. 10 MapReduce Key-Value Stores SQL Implementations

11. 11 Distributed File System MapReduce, e.g. Hadoop Object Store (key-value store), e.g., BigTable, Hbase, Cassandra SQL Implementations, e.g., PIG (relational algebra), HIVE

12. 12  MapReduce (Google) and open-source (Apache) equivalent Hadoop.  Important specialized parallel computing tool.  Cope with compute-node failures.  Avoid restart of the entire job.

13. 13  BigTable (Google), Hbase, Cassandra (Apache), Dynamo (Amazon).  Each row is a key plus values over a flexible set of columns.  Each column component can be a set of values.  Example: Structure of the Web.  Key is a URL.  One column is a set of URL’s – those linked to the page represented by the key.  A second column is the set of URL’s linking to the key.

14. 14  PIG – Yahoo! implementation of relational algebra.  Translates to a sequence of map-reduce operations, using Hadoop.  Hive – open-source (Apache) implementation of a restricted SQL, called QL, over Hadoop.

15. 15  Sawzall – Google implementation of parallel select + aggregation, but using C++.  Dremel – (Google) real restricted SQL, column oriented store.  F1 – (Google) row-oriented, conventional, but massive scale.  Scope – Microsoft implementation of restricted SQL.

16. 16 Formal Definition Implementation Fault-Tolerance Examples: Word-Count, Join

17.  Input: a set of key/value pairs.  User supplies two functions:  map(k,v)  set(k1,v1)  reduce(k1, list(v1))  set(v2)  Technically, the input consists of key-value pairs of some type, but usually only the value is important.  (k1,v1) is an intermediate key/value pair.  Output is the set of (k1,v2) pairs.

18.  MapReduce job =  Map function (inputs -> key-value pairs) +  Reduce function (key and list of values -> outputs).  Map and Reduce Tasks apply Map or Reduce function to (typically) many of their inputs.  Unit of parallelism. 18

19. 19  The Map tasks generate key-value pairs.  Each takes one or more chunks of input from the distributed file system.  The system takes all the key-value pairs from all the Map tasks and sorts them by key.  Then, it forms key-(list-of-associated-values) pairs and passes each key-(value-list) pair to one of the Reduce tasks.

20. 20 Map tasks Reduce tasks Input from DFS Output to DFS “key”-value pairs

21.  We have a large file documents, which are sequences of words.  Count the number of times each distinct word appears in the file.

22. map(key, value): // key: document name; value: text of document FOR (each word w in value) emit(w, 1); reduce(key, value-list): // key: a word; value: an iterator over value-list result = 0; FOR (each count v on value-list) result += v; emit(result);

23. User Program Worker Master Worker fork assign map assign reduce read local write remote read, sort Output File 0 Output File 1 write Chunk 0 Chunk1 Chunk 2 Input Data

24.  Input and final output are stored in the distributed file system.  Scheduler tries to schedule Map tasks “close” to physical storage location of input data – preferably at the same node.  Intermediate results are stored on local file storage of Map and Reduce workers.

25.  Maintain task status: (idle, active, completed).  Idle tasks get scheduled as workers become available.  When a Map task completes, it sends the Master the location and sizes of its intermediate files, one for each Reduce task.  Master pushes location of intermediates to Reduce tasks.  Master pings workers periodically to detect failures.

26.  Rule of thumb: Use several times more Map tasks and Reduce tasks than the number of compute nodes available.  Minimizes skew caused by different tasks taking different amounts of time.  One DFS chunk per Map task is common.

27.  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 communication time by applying Reduce function to values with the same key at the Map task.  Called a combiner.  Works only if Reduce function is commutative and associative.

28.  We need to assure that records with the same intermediate key end up at the same Reduce task.  System uses a default partition function e.g., hash(key) mod R, if there are R Reduce tasks.  Sometimes useful to override.  Example: hash(hostname(URL)) mod R ensures URLs from a host end up at the same Reduce task and therefore appear together in the output.

29. 29  MapReduce is designed to deal with compute nodes failing to execute a task.  Re-executes failed tasks, not whole jobs.  Failure modes: 1.Compute-node failure (e.g., disk crash). 2.Rack communication failure. 3.Software failures, e.g., a task requires Java n; node has Java n-1.

30. 30 1. Matrix-Matrix and Matrix-vector multiplication.  One step of the PageRank iteration was the original application. 2. Relational algebra operations.  We’ll do an example of the join. 3. Many other “embarrassingly parallel” operations.

31.  Map-Reduce job =  Map function (inputs -> key-value pairs) +  Reduce function (key and list of values -> outputs).  Map and Reduce Tasks apply Map or Reduce function to (typically) many of their inputs.  Unit of parallelism.  Mapper = application of the Map function to a single input.  Reducer = application of the Reduce function to a single key-(list of values) pair. 31

32.  Join of R(A,B) with S(B,C) is the set of tuples (a,b,c) such that (a,b) is in R and (b,c) is in S.  Mappers need to send R(a,b) and S(b,c) to the same reducer, so they can be joined there.  Mapper output: key = B-value, value = relation and other component (A or C).  Example: R(1,2) -> (2, (R,1)) S(2,3) -> (2, (S,3)) 32

33. 33 Mapper for R(1,2) R(1,2)(2, (R,1)) Mapper for R(4,2) R(4,2) Mapper for S(2,3) S(2,3) Mapper for S(5,6) S(5,6) (2, (R,4)) (2, (S,3)) (5, (S,6))

34.  There is a reducer for each key.  Every key-value pair generated by any mapper is sent to the reducer for its key. 34

35. 35 Mapper for R(1,2) (2, (R,1)) Mapper for R(4,2) Mapper for S(2,3) Mapper for S(5,6) (2, (R,4)) (2, (S,3)) (5, (S,6)) Reducer for B = 2 Reducer for B = 5

36.  The input to each reducer is organized by the system into a pair:  The key.  The list of values associated with that key. 36

37. 37 Reducer for B = 2 Reducer for B = 5 (2, [(R,1), (R,4), (S,3)]) (5, [(S,6)])

38.  Given key b and a list of values that are either (R, a i ) or (S, c j ), output each triple (a i, b, c j ).  Thus, the number of outputs made by a reducer is the product of the number of R’s on the list and the number of S’s on the list. 38

39. 39 Reducer for B = 2 Reducer for B = 5 (2, [(R,1), (R,4), (S,3)]) (5, [(S,6)]) (1,2,3), (4,2,3)