MapReduce and Hadoop. The reference Big Data stack

Size: px

Start display at page:

Download "MapReduce and Hadoop. The reference Big Data stack"

Estella Park
5 years ago
Views:

Università degli Studi di Roma Tor Vergata Dipartimento di Ingegneria Civile e Ingegneria Informatica MapReduce and Hadoop Corso di Sistemi e Ar

1 Università degli Studi di Roma Tor Vergata Dipartimento di Ingegneria Civile e Ingegneria Informatica MapReduce and Hadoop Corso di Sistemi e Architetture per Big Data A.A. 2017/18 Valeria Cardellini The reference Big Data stack High-level Interfaces Data Processing Data Storage Resource Management Support / Integration Valeria Cardellini - SABD 2017/18 1

2 The Big Data landscape Valeria Cardellini - SABD 2017/18 2 The open-source landscape for Big Data Valeria Cardellini - SABD 2017/18 3

Processing Big Data Some frameworks and platforms to process Big Data Relational DBMS: old fashion, no longer applicable Batch processing: store and process data sets at massive scale (especially

3 Processing Big Data Some frameworks and platforms to process Big Data Relational DBMS: old fashion, no longer applicable Batch processing: store and process data sets at massive scale (especially Volume+Variety) In this lesson: MapReduce and Hadoop Data stream processing: process fast data (in realtime) as data is being generated, without storing Valeria Cardellini - SABD 2017/18 4 MapReduce Valeria Cardellini - SABD 2017/18 5

4 Parallel programming: background Parallel programming Break processing into parts that can be executed concurrently on multiple processors Challenge Identify tasks that can run concurrently and/or groups of data that can be processed concurrently Not all problems can be parallelized! Valeria Cardellini - SABD 2017/18 6 Parallel programming: background Simplest environment for parallel programming No dependency among data Data can be split into equal-size chunks Each process can work on a chunk Master/worker approach Master Initializes array and splits it according to the number of workers Sends each worker the sub-array Receives the results from each worker Worker: Receives a sub-array from master Performs processing Sends results to master Single Program, Multiple Data (SPMD): technique to achieve parallelism The most common style of parallel programming Valeria Cardellini - SABD 2017/18 7

Key idea behind MapReduce: Divide and conquer A feasible approach to tackling large-data problems Partition a large problem into smaller sub-problems Independent sub-problems executed in parallel

5 Key idea behind MapReduce: Divide and conquer A feasible approach to tackling large-data problems Partition a large problem into smaller sub-problems Independent sub-problems executed in parallel Combine intermediate results from each individual worker The workers can be: Threads in a processor core Cores in a multi-core processor Multiple processors in a machine Many machines in a cluster Implementation details of divide and conquer are complex Valeria Cardellini - SABD 2017/18 8 Divide and conquer: how? Decompose the original problem in smaller, parallel tasks Schedule tasks on workers distributed in a cluster, keeping into account: Data locality Resource availability Ensure workers get the data they need Coordinate synchronization among workers Share partial results Handle failures Valeria Cardellini - SABD 2017/18 9

6 Key idea behind MapReduce: Scale out, not up! For data-intensive workloads, a large number of commodity servers is preferred over a small number of high-end servers Cost of super-computers is not linear Datacenter efficiency is a difficult problem to solve, but recent improvements Processing data is quick, I/O is very slow Sharing vs. shared nothing: Sharing: manage a common/global state Shared nothing: independent entities, no common state Sharing is difficult: Synchronization, deadlocks Finite bandwidth to access data from SAN Temporal dependencies are complicated (restarts) Valeria Cardellini - SABD 2017/18 10 MapReduce Programming model for processing huge amounts of data sets over thousands of servers Originally proposed by Google in 2004: MapReduce: simplified data processing on large clusters Based on a shared nothing approach Also an associated implementation (framework) of the distributed system that runs the corresponding programs Some examples of applications for Google: Web indexing Reverse Web-link graph Distributed sort Web access statistics Valeria Cardellini - SABD 2017/18 11

MapReduce: programmer view MapReduce hides system-level details Key idea: separate the what from the how MapReduce abstracts away the distributed part of the system Such

Cardellini - SABD 2017/18 12 Typical Big Data problem Iterate over a large number of records Extract something of interest from each record Shuffle and sort intermediate

7 MapReduce: programmer view MapReduce hides system-level details Key idea: separate the what from the how MapReduce abstracts away the distributed part of the system Such details are handled by the framework Programmers get simple API Don t have to worry about handling Parallelization Data distribution Load balancing Fault tolerance Valeria Cardellini - SABD 2017/18 12 Typical Big Data problem Iterate over a large number of records Extract something of interest from each record Shuffle and sort intermediate results Aggregate intermediate results Generate final output Key idea: provide a functional abstraction of the two Map and Reduce operations Valeria Cardellini - SABD 2017/18 13

8 MapReduce: model Processing occurs in two phases: Map and Reduce Functional programming roots (e.g., Lisp) Map and Reduce are defined by the programmer Input and output: sets of key-value pairs Programmers specify two functions: map and reduce map(k 1, v 1 ) [(k 2, v 2 )] reduce(k 2, [v 2 ]) [(k 3, v 3 )] (k, v) denotes a (key, value) pair [ ] denotes a list Keys do not have to be unique: different pairs can have the same key Normally the keys k 1 of input elements are not relevant Valeria Cardellini - SABD 2017/18 14 Map Execute a function on a set of key-value pairs (input shard) to create a new list of key-value pairs map (in_key, in_value) list(out_key, intermediate_value) Map calls are distributed across machines by automatically partitioning the input data into shards Parallelism is achieved as keys can be processed by different machines MapReduce library groups together all intermediate values associated with the same intermediate key and passes them to the Reduce function Valeria Cardellini - SABD 2017/18 15

Reduce Combine values in sets to create a new value reduce (out_key, list(intermediate_value))

is achieved as reducers operating on different keys can be executed simultaneously Valeria Cardellini

numbers from 1 to n) in MapReduce fashion Map function: map square [1,2,3,4] returns [1,4,9,16]

9 Reduce Combine values in sets to create a new value reduce (out_key, list(intermediate_value)) list(out_key, out_value) - The key in output is often identical to the key in the input - Parallelism is achieved as reducers operating on different keys can be executed simultaneously Valeria Cardellini - SABD 2017/18 16 Your first MapReduce example (in Lisp) Example: sum-of-squares (sum the square of numbers from 1 to n) in MapReduce fashion Map function: map square [1,2,3,4] returns [1,4,9,16] Reduce function: reduce [1,4,9,16] returns 30 (the sum of the square elements) Valeria Cardellini - SABD 2017/18 17

10 MapReduce program A MapReduce program, referred to as a job, consists of: Code for Map and Reduce packaged together Configuration parameters (where the input lies, where the output should be stored) Input data set, stored on the underlying distributed file system The input will not fit on a single computer s disk Each MapReduce job is divided by the system into smaller units called tasks Map tasks or mappers Reduce tasks or reducers All mappers need to finish before reducers can begin The output of MapReduce job is also stored on the underlying distributed file system A MapReduce program may consist of many rounds of different map and reduce functions Valeria Cardellini - SABD 2017/18 18 MapReduce computation 1. Some number of Map tasks each are given one or more chunks of data from a distributed file system 2. These Map tasks turn the chunk into a sequence of key-value pairs The way key-value pairs are produced from the input data is determined by the code written by the user for the Map function 3. The key-value pairs from each Map task are collected by a master controller and sorted by key 4. The keys are divided among all the Reduce tasks, so all keyvalue pairs with the same key wind up at the same Reduce task 5. The Reduce tasks work on one key at a time, and combine all the values associated with that key in some way The manner of combination of values is determined by the code written by the user for the Reduce function 6. Output key-value pairs from each reducer are written persistently back onto the distributed file system 7. The output ends up in r files, where r is the number of reducers Such output may be the input to a subsequent MapReduce phase Valeria Cardellini - SABD 2017/18 19

Where the magic happens Implicit between the map and reduce phases is a distributed group by operation on intermediate keys, called shuffle and sort Transfer mappers output to reducers, merging and

11 Where the magic happens Implicit between the map and reduce phases is a distributed group by operation on intermediate keys, called shuffle and sort Transfer mappers output to reducers, merging and sorting the output Intermediate data arrive at every reducer sorted by key Intermediate keys are transient They are not stored on the distributed file system They are spilled to the local disk of each machine in the cluster Input Splits Intermediate Outputs Final Outputs (k, v) Pairs Map Function (k, v ) Pairs Reduce Function (k, v ) Pairs Valeria Cardellini - SABD 2017/18 20 MapReduce computation: the complete picture Valeria Cardellini - SABD 2017/18 21

A simplified view of MapReduce: example Mappers are applied to all input key-value pairs, to generate an arbitrary number of intermediate pairs Reducers are applied to all intermediate values

12 A simplified view of MapReduce: example Mappers are applied to all input key-value pairs, to generate an arbitrary number of intermediate pairs Reducers are applied to all intermediate values associated with the same intermediate key Between the map and reduce phase lies a barrier that involves a large distributed sort and group by Valeria Cardellini - SABD 2017/18 22 Hello World in MapReduce: WordCount Problem: count the number of occurrences for each word in a large collection of documents Input: repository of documents, each document is an element Map: read a document and emit a sequence of key-value pairs where: Keys are words of the documents and values are equal to 1: (w 1, 1), (w 2, 1),, (w n, 1) Shuffle and sort: group by key and generates pairs of the form (w 1, [1, 1,, 1]),..., (w n, [1, 1,, 1]) Reduce: add up all the values and emits (w 1, k),, (w n, l) Output: (w,m) pairs where: w is a word that appears at least once among all the input documents and m is the total number of occurrences of w among all those documents Valeria Cardellini - SABD 2017/18 23

WordCount in practice Map Shuffle Reduce Valeria Cardellini - SABD 2017/18 24 WordCount: Map Map emits each word in the document with an associated value equal to 1

13 WordCount in practice Map Shuffle Reduce Valeria Cardellini - SABD 2017/18 24 WordCount: Map Map emits each word in the document with an associated value equal to 1 Map(String key, String value): // key: document name // value: document contents for each word w in value: EmitIntermediate(w, 1 ); Valeria Cardellini - SABD 2017/18 25

14 WordCount: Reduce Reduce adds up all the 1 emitted for a given word Reduce(String key, Iterator values): // key: a word // values: a list of counts int result=0 for each v in values: result += ParseInt(v) Emit(AsString(result)) This is pseudo-code; for the complete code of the example see the MapReduce paper Valeria Cardellini - SABD 2017/18 26 Example: WordLengthCount Problem: count how many words of certain lengths exist in a collection of documents Input: a repository of documents, each document is an element Map: read a document and emit a sequence of key-value pairs where the key is the length of a word and the value is the word itself: (i, w1),, (j, wn) Shuffle and sort: group by key and generate pairs of the form (1, [w1,, wk]),, (n, [wr,, ws]) Reduce: count the number of words in each list and emit: (1, l),, (p, m) Output: (l,n) pairs, where l is a length and n is the total number of words of length l in the input documents Valeria Cardellini - SABD 2017/18 27

15 Example: matrix-vector multiplication Sparse matrix A = [a ij ] size n x n Vector x = [x j ] size n x 1 Problem: matrix-vector multiplication y = Ax where y i = Σ j=1 n a ij x j Used in the PageRank algorithm Let us assume that x can fit into the mapper memory Map: apply to ((i, j), a ij ) and produce key-value pair (i, a ij x j ) Reduce: receive (i, [a i1 x 1,..., a in x n ]) and sum all values of the list associated with a given key i, i.e., y i = Σ j=1 n a ij x j. The result will be a pair (i, y i ) Valeria Cardellini - SABD 2017/18 28 Example: matrix-vector multiplication What if vector x does not fit in mapper s memory? Solution: - Split x in horizontal stripes that fit in memory - Split matrix A accordingly in vertical stripes - Stripes of A do not need to fit in memory y A x Each mapper is assigned a matrix stripe; it also gets the corresponding vector stripe Map and Reduce functions are as before The best strategy is to partition A into square blocks, rather than stripes Valeria Cardellini - SABD 2017/18 29

Combine When the Reduce function is associative and commutative, we can push some of what the reducers do to the preceding Map tasks Associative and commutative: values to be combined can be combined

16 Combine When the Reduce function is associative and commutative, we can push some of what the reducers do to the preceding Map tasks Associative and commutative: values to be combined can be combined in any order, with the same result E.g., addition in Reduce of WordCount In this case we also apply a combiner to Map combine (k2, [v2]) [(k3, v3)] In many cases the same function can be used for combining as the final reduction Shuffle and sort is still necessary! Advantages: Reduce amount of intermediate data Reduce network traffic Valeria Cardellini - SABD 2017/18 30 Combine Valeria Cardellini - SABD 2017/18 31

17 MapReduce: execution overview Master-worker architecture The master controller process coordinates the map and reduce tasks running the Map Reduce job Can rerun tasks on a failed compute node The worker processes execute the map and reduce tasks Valeria Cardellini - SABD 2017/18 32 MapReduce: execution overview Valeria Cardellini - SABD 2017/18 33

will have to be redone, even if they had completed Computer node hosting a reduce worker fails Reschedule

18 Coping with failures Worst scenario: compute node hosting the master fails The entire MapReduce job must be restarted Compute node hosting a map worker fails All the Map tasks that were assigned to this Worker will have to be redone, even if they had completed Computer node hosting a reduce worker fails Reschedule on another reduce worker Valeria Cardellini - SABD 2017/18 34 Apache Hadoop Valeria Cardellini - SABD 2017/18 35

What is Apache Hadoop? Open-source software framework for reliable, scalable, distributed data-intensive computing Originally developed by Yahoo!

19 What is Apache Hadoop? Open-source software framework for reliable, scalable, distributed data-intensive computing Originally developed by Yahoo! Goal: storage and processing of data-sets at massive scale Infrastructure: clusters of commodity hardware Core components: HDFS: Hadoop Distributed File System Hadoop YARN Hadoop MapReduce Includes a number of related projects Among which Apache Pig, Apache Hive, Apache HBase Used in production by Facebook, IBM, Linkedin, Twitter, Yahoo! and many others Valeria Cardellini - SABD 2017/18 36 Hadoop runs on clusters Compute nodes are stored on racks 8-64 compute nodes on a rack There can be many racks of compute nodes The nodes on a single rack are connected by a network, typically gigabit Ethernet Racks are connected by another level of network or a switch The bandwidth of intra-rack communication is usually much greater than that of inter-rack communication Compute nodes can fail! Solution: Files are stored redundantly Computations are divided into tasks Valeria Cardellini - SABD 2017/18 37

20 Hadoop runs on commodity hardware Pros You are not tied to expensive, proprietary offerings from a single vendor You can choose standardized, commonly available hardware from any of a large range of vendors to build your cluster Scale out, not up Cons Significant failure rates Solution: replication! Valeria Cardellini - SABD 2017/18 38 Commodity hardware A typical specification of commodity hardware Processor: Intel Xeon E cores GHz Memory: GB ECC RAM Storage: 4 1TB HDD Network: 1-10 Gb Ethernet Yahoo! has a huge installation of Hadoop: In 2014: more than 100,000 CPUs in > 40,000 computers running Hadoop Used to support research for Ad Systems and Web Search Also used to do scaling tests to support development of Apache Hadoop on larger clusters In 2017: 36 different clusters spread across YARN, Apache HBase, and Apache Storm deployments For some example, see Valeria Cardellini - SABD 2017/18 39

on large clusters (> 1000 nodes) of commodity hardware in a reliable,

21 Hadoop core HDFS ü Already examined Hadoop YARN Cluster resource management Hadoop MapReduce System to write applications which process large data sets in parallel on large clusters (> 1000 nodes) of commodity hardware in a reliable, fault-tolerant manner Valeria Cardellini - SABD 2017/18 40 Valeria Cardellini - SABD 2017/18 41

22 Hadoop core: MapReduce Programming paradigm for expressing distributed computations over multiple servers The powerhouse behind most of today s big data processing Also used in other MPP environments and NoSQL databases (e.g., Vertica and MongoDB) Improving Hadoop usage Improving programmability: Pig and Hive Improving data access: HBase, Sqoop and Flume Valeria Cardellini - SABD 2017/18 42 We will study YARN later Global ResourceManager (RM) A set of per-application ApplicationMasters (AMs) A set of NodeManagers (NMs) Valeria Cardellini - SABD 2017/18 43

(data locality optimization) So to not use cluster bandwidth

23 YARN job execution Valeria Cardellini - SABD 2017/18 44 Data locality optimization Hadoop tries to run each map task on a datalocal node (data locality optimization) So to not use cluster bandwidth Otherwise rack-local Off-rack as last choice Valeria Cardellini - SABD 2017/18 45

multiple Reduce tasks When there are multiple reducers,

24 MapReduce data flow: single Reduce task Valeria Cardellini - SABD 2017/18 46 MapReduce data flow: multiple Reduce tasks When there are multiple reducers, the map tasks partition their output Valeria Cardellini - SABD 2017/18 47

Partitioner When there are multiple reducers, the map tasks partition their output One

intermediate keys and values Records for any given key are all in a single partition Default

uses a hash function on the key to determine which bucket (i.e., reducer) Partitioning can be

25 Partitioner When there are multiple reducers, the map tasks partition their output One partition for each reduce task Goal: to determine which reducer will receive which intermediate keys and values Records for any given key are all in a single partition Default partitioner uses a hash function on the key to determine which bucket (i.e., reducer) Partitioning can be also controlled by a user-defined function Requires to implement a custom partitioner Valeria Cardellini - SABD 2017/18 48 MapReduce data flow: no Reduce task Valeria Cardellini - SABD 2017/18 49

26 Combiner Many MapReduce jobs are limited by the cluster bandwidth How to minimize the data transferred between map and reduce tasks? Use Combine task A combiner is an optional localized reducer that applies a userprovided method to combine mapper output Performs data aggregation on intermediate data of the same key for the Map task s output before transmitting the result to the Reduce task Takes each key-value pair from the Map task, processes it, and produces the output as key-value collection pairs The Reduce task is still needed to process records with the same key from different maps Valeria Cardellini - SABD 2017/18 50 Programming languages for Hadoop Java is the default programming language In Java you write a program with at least three parts: 1. A Main method which configures the job, and lauches it Set number of reducers Set mapper and reducer classes Set partitioner Set other Hadoop configurations 2. A Mapper class Takes (k,v) inputs, writes (k,v) outputs 3. A Reducer class Takes k, Iterator[v] inputs, and writes (k,v) outputs Valeria Cardellini - SABD 2017/18 51

27 Programming languages for Hadoop Hadoop Streaming API to write Map and Reduce functions in programming languages other than Java Uses Unix standard streams as interface between Hadoop and MapReduce program Allows to use any language (e.g., Python) that can read standard input (stdin) and write to standard output (stdout) to write the MapReduce program Standard input is stream data going into a program; it is from the keyboard Standard output is the stream where a program writes its output data; it is the text terminal Reducer interface for streaming is different than in Java: instead of receiving reduce(k, Iterator[v]), your script is sent one line per value, including the key Valeria Cardellini - SABD 2017/18 52 WordCount on Hadoop Let s analyze the WordCount code The code in Java The same code in Python Use Hadoop Streaming API for passing data between the Map and Reduce code via standard input and standard output See Michael Noll s tutorial Valeria Cardellini - SABD 2017/18 53

28 WordCount in Java import java.io.ioexception; import java.util.stringtokenizer; import org.apache.hadoop.conf.configuration; import org.apache.hadoop.fs.path; import org.apache.hadoop.io.intwritable; import org.apache.hadoop.io.text; import org.apache.hadoop.mapreduce.job; import org.apache.hadoop.mapreduce.mapper; import org.apache.hadoop.mapreduce.reducer; import org.apache.hadoop.mapreduce.lib.input.fileinputformat; import org.apache.hadoop.mapreduce.lib.output.fileoutputformat; Valeria Cardellini - SABD 2017/18 54 public class WordCount { WordCount in Java (2) map method processes one line at a time, splits the line into tokens separated by whitespaces, via StringTokenizer, and emits a key-value pair <word, 1> public static class TokenizerMapper extends Mapper<Object, Text, Text, IntWritable>{ private final static IntWritable one = new IntWritable(1); private Text word = new Text(); public void map(object key, Text value, Context context ) throws IOException, InterruptedException { StringTokenizer itr = new StringTokenizer(value.toString()); while (itr.hasmoretokens()) { word.set(itr.nexttoken()); context.write(word, one); } } } Valeria Cardellini - SABD 2017/18 55

WordCount in Java (3) reduce method sums up the values, which are the occurrence counts for each key public static class IntSumReducer extends Reducer<Text,IntWritable,Text,IntWritable> { private

29 WordCount in Java (3) reduce method sums up the values, which are the occurrence counts for each key public static class IntSumReducer extends Reducer<Text,IntWritable,Text,IntWritable> { private IntWritable result = new IntWritable(); } public void reduce(text key, Iterable<IntWritable> values, } int sum = 0; Context context ) throws IOException, InterruptedException { for (IntWritable val : values) { sum += val.get(); } result.set(sum); context.write(key, result); Valeria Cardellini - SABD 2017/18 56 } WordCount in Java (4) public static void main(string[] args) throws Exception { Configuration conf = new Configuration(); } main method specifies various facets of the job Job job = Job.getInstance(conf, "word count"); job.setjarbyclass(wordcount.class); job.setmapperclass(tokenizermapper.class); job.setcombinerclass(intsumreducer.class); job.setreducerclass(intsumreducer.class); job.setoutputkeyclass(text.class); job.setoutputvalueclass(intwritable.class); FileInputFormat.addInputPath(job, new Path(args[0])); FileOutputFormat.setOutputPath(job, new Path(args[1])); System.exit(job.waitForCompletion(true)? 0 : 1); the output of each map is passed through the local combiner (same as the Reducer) for local aggregation, after being sorted on the keys Valeria Cardellini - SABD 2017/18 57

30 WordCount in Java (5) Input files Hello World Bye World Hello Hadoop Goodbye Hadoop Output file Bye 1 Goodbye 1 Hadoop 2 Hello 2 World 2 Valeria Cardellini - SABD 2017/18 58 WordCount in Java (6) 1st mapper output < Hello, 1> < World, 1> < Bye, 1> < World, 1> After local combiner < Bye, 1> < Hello, 1> < World, 2> Valeria Cardellini - SABD 2017/18 59

31 WordCount in Java (7) 2nd mapper output After local combiner < Hello, 1> < Hadoop, 1> < Goodbye, 1> < Hadoop, 1> < Goodbye, 1> < Hadoop, 2> < Hello, 1> Reducer output < Bye, 1> < Goodbye, 1> < Hadoop, 2> < Hello, 2> < World, 2> Valeria Cardellini - SABD 2017/18 60 Hadoop installation Some straightforward alternatives Single-node installation Cloudera CDH, Hortonworks HDP, MapR distributions MapR Integrated Hadoop-based stack containing all the components needed for production, tested and packaged to work together Include at least Hadoop, YARN, HDFS (not in MapR), HBase, Hive, Pig, Sqoop, Flume, Zookeeper, Kafka, Spark Own distributed Posix file system (MapR-FS) rather than HDFS Valeria Cardellini - SABD 2017/18 61

Hadoop installation Cloudera CDH See QuickStarts http://bit.

production Valeria Cardellini - SABD 2017/18 62 Hadoop installation Hortonworks HDP Hortonworks Sandbox on a VM Requires either VirtualBox, VMWare or

32 Hadoop installation Cloudera CDH See QuickStarts Single-node cluster: requires to have previously installed an hypervisor (VMware, KVM, or VirtualBox) or Docker QuickStarts not for use in production Valeria Cardellini - SABD 2017/18 62 Hadoop installation Hortonworks HDP Hortonworks Sandbox on a VM Requires either VirtualBox, VMWare or Docker Image file sizes: GB Includes Ambari: open source management platform for Hadoop Distributions include security frameworks (Sentri, Ranger, Knobs, ) to handle access control, security policies, Valeria Cardellini - SABD 2017/18 63

33 Hadoop configuration Tuning Hadoop clusters for good performance is somehow magic HW and SW systems (including OS), e.g.: Optimal number of disks that maximizes I/O bandwidth BIOS parameters File system type (on Linux, EXT4 is better) Open file handles Optimal HDFS block size JVM settings for Java heap usage and garbage collections See for some detailed hints Valeria Cardellini - SABD 2017/18 64 Hadoop configuration There are also Hadoop-specific parameters that can be tuned for performance Number of mappers Number of reducers Many other map-side and reduce-side tuning parameters Valeria Cardellini - SABD 2017/18 65

34 How many mappers Number of mappers Driven by the number of blocks in the input files You can adjust the HDFS block size to adjust the number of maps Good level of parallelism: mappers per-node, but up to 300 mappers for very CPU-light map tasks Task setup takes a while, so it is best if the maps take at least a minute to execute Valeria Cardellini - SABD 2017/18 66 How many reducers Number of reducers Can be user-defined (default is 1) Use Job.setNumReduceTasks(int) The right number of reduces seems to be 0.95 or 1.75 multiplied by (<no. of nodes> * <no. of maximum containers per node>) 0.95: all of the reduces can launch immediately and start transferring map outputs as the maps finish 1.75: the faster nodes will finish their first round of reduces and launch a second wave of reduces doing a much better job of load balancing Can be set to zero if no reduction is desired No sorting of map-outputs before writing them to output file See Valeria Cardellini - SABD 2017/18 67

Some tips for performance To handle massive I/O and save bandwidth - Compress input data, map output data, reduce output data To address massive I/O in partition and sort phases Each map task has a

35 Some tips for performance To handle massive I/O and save bandwidth - Compress input data, map output data, reduce output data To address massive I/O in partition and sort phases Each map task has a circular buffer memory to which it writes the output; when the buffer is full, contents are written ( spilled ) to disk. Avoid that records are spilled more than once How? Adjust spill records and sorting buffer To address massive network traffic caused by large map output - Compress map output - Implement a combiner to reduce the amount of data passing through the shuffle See Valeria Cardellini - SABD 2017/18 68 Hadoop in the Cloud Pros: Gain Cloud scalability and elasticity Do not need to manage and provision the infrastructure and the platform Main challenges: Move data to the Cloud Latency is not zero (because of speed of light)! Minor issue: network bandwidth Data security and privacy Valeria Cardellini - SABD 2017/18 69

Amazon Elastic MapReduce (EMR) Distribute the computational work across a cluster of virtual servers running in AWS cloud (EC2 instances) Not only Hadoop: also Hive, Pig, Spark and Presto Usually not

36 Amazon Elastic MapReduce (EMR) Distribute the computational work across a cluster of virtual servers running in AWS cloud (EC2 instances) Not only Hadoop: also Hive, Pig, Spark and Presto Usually not the latest released version Input and output: Amazon S3, DynamoDB, Elastic File System (EFS) Access through AWS Management Console, command line Valeria Cardellini - SABD 2017/18 70 Create EMR cluster Steps: 1. Provide cluster name 2. Select EMR release and applications to install 3. Select instance type (including spot instances) 4. Select number of instances 5. Select EC2 key-pair to connect securely to the cluster See Valeria Cardellini - SABD 2017/18 71

37 EMR cluster details You can still tune some parameters for performance - Some EC2 parameters (heap size used by Hadoop and Yarn ) - Hadoop parameters (e.g., memory for map and reduce JVMs) Valeria Cardellini - SABD 2017/18 72 Google Cloud Dataproc Distribute the computational work across a cluster of virtual servers running in the Google Cloud Platform Not only Hadoop (plus Hive and Pig), also Spark Input and output from other Google services: Cloud Storage, Bigtable Access through REST API, Cloud SDK, and Cloud Dataproc UI Fine-grain pay-per-use (seconds) Valeria Cardellini - SABD 2017/18 73

Create Cloud Datproc cluster Valeria Cardellini - SABD 2017/18 74 References Dean and Ghemawat, "MapReduce: simplified data processing on large clusters", OSDI 2004. http://static.googleusercontent.

38 Create Cloud Datproc cluster Valeria Cardellini - SABD 2017/18 74 References Dean and Ghemawat, "MapReduce: simplified data processing on large clusters", OSDI it//archive/mapreduce-osdi04.pdf White, Hadoop: The Definitive Guide, 4 th edition, O Reilly, Miner and Shook, MapReduce Design Patterns, O Reilly, Leskovec, Rajaraman, and Ullman, Mining of Massive Datasets, chapter 2, Valeria Cardellini - SABD 2017/18 75

MapReduce and Hadoop

Università degli Studi di Roma Tor Vergata MapReduce and Hadoop Corso di Sistemi e Architetture per Big Data A.A. 2016/17 Valeria Cardellini The reference Big Data stack High-level Interfaces Data Processing