Implementing an iterative computation by running a separate Hadoop job for every iteration is usually not very efficient (although it is fault tolerant).

If we have enough machines that all input data fits into memory, we can implement iterative computation like this:

1. start several machines, each reads a portion of input data
2. when all the machines are ready, for each iteration
   1. locally compute as much as is possible
   2. communicate with other machines and combine the results to complete this iteration

There are many ways how the machines communicate with one another to finish an iteration. On of the possible implementations is AllReduce operation.

An AllReduce operation does the following:

• every machine provides a value
• values from all the machines are combined using specified operation (e.g., addition, minimum, maximum) into one resulting value
• the resulting value is distributed to all the machines

It is possible for each machine to provide more than value. In that case, all machines must provide the same number of values, the corresponding values get reduced in parallel and every machine gets the result of all the reductions.

An implementation of AllReduce operation is provided in our Hadoop framework. The implementation is in theory independent of Hadoop (it could be implemented using e.g. an SGE array job), but we use Hadoop framework as it provides features we would have to implement ourselves.

A Hadoop AllReduce job is implemented using a mapper only. The mapper should read the input data during map operation without providing any output. After all input data is read, cooperate function is called. As the name suggests, it is executed by all the machines in a synchronized fashion. In this method, the following allReduce methods can be called:

• double allReduce(Context, double value, int reduce_op) – each machine provides one value, which is reduced across all the machines and the result is returned to all machines.
• void allReduce(Context, double[] values, int reduce_op) – each machine provides multiple values. The reduced values are stored in the original array in every machine.
• void allReduce(Context, double[][] values, int reduce_op) – each machine provides multiple values. The reduced values are stored in the original array in every machine.

The reduce_op can currently be one on:

• REDUCE_ADD
• REDUCE_MIN
• REDUCE_MAX

It is crucial that all the mappers run simultaneously. This can be achieved using the /net/projects/hadoop/bin/compute-splitsize script: for given Hadoop input and requested number of mappers, it computes the appropriate splitsize.

When the computation finishes, only one of the mappers should print the results, as all of them have the same results. For simplicity, the cooperate method has boolean writeResults argument, which is set in exactly one mapper.

This example reads the keys of /net/projects/hadoop/examples/inputs/numbers-small, computes the sum of all the keys and print it:

Sum.java

import java.io.IOException;
 
import org.apache.hadoop.conf.*;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.*;
import org.apache.hadoop.mapreduce.*;
import org.apache.hadoop.mapreduce.lib.allreduce.*;
import org.apache.hadoop.mapreduce.lib.input.*;
import org.apache.hadoop.mapreduce.lib.output.*;
import org.apache.hadoop.util.*;
 
public class Sum extends Configured implements Tool {
  public static class TheMapper extends AllReduceMapper<Text, Text, DoubleWritable, NullWritable>{
    int[] points = new int[64];
    int points_num = 0;
 
    public void map(Text key, Text value, Context context) throws IOException, InterruptedException {
      if (points_num == points.length) {
        int[] new_points = new int[2*points_num];
        System.arraycopy(points, 0, new_points, 0, points_num);
        points = new_points;
      }
 
      points[points_num++] = Integer.parseInt(key.toString());
    }
 
    public void cooperate(Context context, boolean writeResults) throws IOException, InterruptedException {
      double sum = 0;
      for (int i = 0; i < points_num; i++) sum += points[i];
 
      double total_sum = allReduce(context, sum, REDUCE_ADD);
 
      if (writeResults) context.write(new DoubleWritable(total_sum), NullWritable.get());
    }
  }
 
  // Job configuration
  public int run(String[] args) throws Exception {
    if (args.length < 2) {
      System.err.printf("Usage: %s.jar in-path out-path", this.getClass().getName());
      return 1;
    }
 
    Job job = new Job(getConf(), this.getClass().getName());
 
    job.setJarByClass(this.getClass());
    job.setMapperClass(TheMapper.class);
    AllReduce.init(job);
    job.setOutputKeyClass(DoubleWritable.class);
    job.setOutputValueClass(NullWritable.class);
 
    job.setInputFormatClass(KeyValueTextInputFormat.class);
 
    FileInputFormat.addInputPath(job, new Path(args[0]));
    FileOutputFormat.setOutputPath(job, new Path(args[1]));
 
    return job.waitForCompletion(true) ? 0 : 1;
  }
 
  // Main method
  public static void main(String[] args) throws Exception {
    int res = ToolRunner.run(new Sum(), args);
 
    System.exit(res);
  }
}

You can run the example locally using:

wget --no-check-certificate 'https://wiki.ufal.ms.mff.cuni.cz/_export/code/courses:mapreduce-tutorial:step-31?codeblock=0' -O Sum.java
make -f /net/projects/hadoop/java/Makefile Sum.jar
rm -rf step-31-out; /net/projects/hadoop/bin/hadoop Sum.jar /net/projects/hadoop/examples/inputs/numbers-small step-31-out
less step-31-out/part-*

To run on a cluster with C machines using C mappers:

rm -rf step-31-out; /net/projects/hadoop/bin/hadoop Sum.jar -c C `/net/projects/hadoop/bin/compute-splitsize /net/projects/hadoop/examples/inputs/numbers-small C` /net/projects/hadoop/examples/inputs/numbers-small step-31-out
less step-31-out/part-*

Implement an AllReduce job on /net/projects/hadoop/examples/inputs/numbers-small, which computes

• number of keys
• mean of the keys
• variance of the keys
• minimum of the keys
• maximum of the keys

You can download the template Statistics.java and execute it using:

wget --no-check-certificate 'https://wiki.ufal.ms.mff.cuni.cz/_media/courses:mapreduce-tutorial:step-31-exercise1.txt' -O Statistics.java
# NOW VIEW THE FILE
# $EDITOR Statistics.java
make -f /net/projects/hadoop/java/Makefile Statistics.jar
rm -rf step-31-out; /net/projects/hadoop/bin/hadoop Statistics.jar -c C `/net/projects/hadoop/bin/compute-splitsize /net/projects/hadoop/examples/inputs/numbers-small C` /net/projects/hadoop/examples/inputs/numbers-small step-31-out
less step-31-out/part-*

Implement an AllReduce job on /net/projects/hadoop/examples/inputs/numbers-small, which computes median of the input data. You can use the following iterative algorithm:

• At the beginning, set min₁ = Integer.MIN_VALUE, max₁ = Integer.MAX_VALUE, index_to_find = number_of_input_data / 2.
• In step i, do the following:
  1. Consider only input keys in range <min_i, max_i>.
  2. Compute split = ceiling of mean of the keys.
  3. If the index_to_find is in range <1+number of keys less than split, number of keys less or equal to split>, then split is median.
  4. Else, if index_to_find is at most the number of keys less than split, set max_i+1 = split-1.
  5. Else, set min_i+1 = split+1 and subtract from index_to_find the number of keys less or equal to split.

You can download the template Median.java and execute it using:

wget --no-check-certificate 'https://wiki.ufal.ms.mff.cuni.cz/_media/courses:mapreduce-tutorial:step-31-exercise2.txt' -O Median.java
# NOW VIEW THE FILE
# $EDITOR Median.java
make -f /net/projects/hadoop/java/Makefile Median.jar
rm -rf step-31-out; /net/projects/hadoop/bin/hadoop Median.jar -c C `/net/projects/hadoop/bin/compute-splitsize /net/projects/hadoop/examples/inputs/numbers-small C` /net/projects/hadoop/examples/inputs/numbers-small step-31-out
less step-31-out/part-*

Solution: Median.java.

Implement an AllReduce job on /net/projects/hadoop/examples/inputs/points-small, which implements the K-means clustering algorithm. See K-means clustering exercise for description of input data.

You can download the template KMeans.java. This template uses two Hadoop properties:

• clusters.num – number of clusters
• clusters.file – file where to read the initial clusters from

You can download and compile it using:

wget --no-check-certificate 'https://wiki.ufal.ms.mff.cuni.cz/_media/courses:mapreduce-tutorial:step-31-exercise3.txt' -O KMeans.java
# NOW VIEW THE FILE
# $EDITOR KMeans.java
make -f /net/projects/hadoop/java/Makefile KMeans.jar

You can run it using specified number of machines on the following input data:

• /net/projects/hadoop/examples/inputs/points-small:

M=machines; K=50; INPUT=/net/projects/hadoop/examples/inputs/points-small/points.txt
rm -rf step-31-out; /net/projects/hadoop/bin/hadoop KMeans.jar -Dclusters.num=$K -Dclusters.file=$INPUT [-jt jobtracker | -c $M] `/net/projects/hadoop/bin/compute-splitsize $INPUT $M` $INPUT step-31-out

• /net/projects/hadoop/examples/inputs/points-medium:

M=machines; K=100; INPUT=/net/projects/hadoop/examples/inputs/points-medium/points.txt
rm -rf step-31-out; /net/projects/hadoop/bin/hadoop KMeans.jar -Dclusters.num=$K -Dclusters.file=$INPUT [-jt jobtracker | -c $M] `/net/projects/hadoop/bin/compute-splitsize $INPUT $M` $INPUT step-31-out

• /net/projects/hadoop/examples/inputs/points-large:

M=machines; K=200; INPUT=/net/projects/hadoop/examples/inputs/points-large/points.txt
rm -rf step-31-out; /net/projects/hadoop/bin/hadoop KMeans.jar -Dclusters.num=$K -Dclusters.file=$INPUT [-jt jobtracker | -c $M] `/net/projects/hadoop/bin/compute-splitsize $INPUT $M` $INPUT step-31-out

Solution: KMeans.java.