近日，需要用ParMetis對大圖資料進行分割槽，其輸入是無向圖（鄰接表形式）且按照頂點ID排序，於是想到用Hadoop中的TeraSort演算法對無向圖進行排序。但Hadoop自帶TeraSort演算法是按照每行資料的前兩個字元排序的，不能滿足要求。

由於圖一般都是用鄰接表的形式儲存，改進的TeraSort演算法就是按照頂點ID進行排序，支援有向圖和無向圖，邊上可附加權值。下面以無向圖為例講述資料的輸入格式。對於下圖，

其輸入資料格式如下，以 \t  間隔，每行第一列為頂點ID。

1 2 3 4

3 1 2

2 1  3 4

4 1 2

擴充套件：只要每行的資料格式滿足：key+“\t”+value，其中key為int或long型，value型別任意。修改的TeraSort演算法就能按照key來對每一行進行排序。

修改方法：

1. 由於輸入格式變化，故首先修改TeraRecordReader類，主要是 boolean next(LongWritable key, Text value)方法，修改如何解析每行資料。程式碼如下：

    public boolean next(LongWritable key, Text value) throws IOException {
      if (in.next(junk, line)) {
        String[] temp=line.toString().split("\t");
        key.set(Long.parseLong(temp[0]));
        if(temp.length!=1) {
          value.set(line.toString().substring(temp[0].length()+1));
        } else {
          value.set("");
        }
        return true;
      } else {
        return false;
      }
    }

2.  JobClient端的資料取樣、排序、獲取分割點等都和原TeraSort演算法類似，注意把key的型別由Text修改為LongWritable型別。

3. 在原TeraSort演算法中，每個map task首先從分散式快取中讀取分割點，然後根據分割點簡歷2-Trie樹。map task從split中依次讀入每條資料，通過Trie樹查詢每條記錄所對應的reduce task編號。 

現在由於是Long型，則不需要構建Trie樹。已知分割點是儲存在splitPoints[]陣列中，按照如下公式計算reduce number，其中length等於splitPoints.length

假設reduce task數目為4（由使用者設定），分割點為34、67、97。則分割點和reduce task編號的對映關係如下：

可以看到小於34的對應第0個reduce task，34和67之間的對應第一個reduce task，67和97之間的key對應第2個reduce task，大於等於97的則對應於第3個reduce task。

主要修改int getPartition(LongWritable key, Text value, int numPartitions)方法，如下：

    @Override
    public int getPartition(LongWritable key, Text value, int numPartitions) {
      if(key.get()<splitPoints[0].get()) {
        return 0;
      }
      for(int i=0;i<splitPoints.length-1;i++) {
        if(key.get()>=splitPoints[i].get() && key.get()<splitPoints[i+1].get()) {
           return i+1;
        }
      }
      return splitPoints.length;
    }

4. 棄用TeraOutputFormat，採用預設的輸出格式就行。

job.setOutputFormat(TextOutputFormat.class);

5. 打成Jar包(TeraSort.jar)，在叢集上執行即可。例如：hadoop jar TeraSort.jar TeraSortTest output 4 

注意輸入引數為：<input> <output> <reduce number>  。與原TeraSort中用 -D mapred.reduce.tasks=value 不同，此處讓使用者明確指定reduce tash的數目。防止使用者忘寫的話，原TeraSort就啟動一個reduce task，那麼整個TeraSort演算法就失去意義！ 執行結果如下：

6. 上述排好序的檔案，依次存在output資料夾下的：part-00000、part-00001、part-00002、part-00003。

 使用 hadoop fs -getmerge output output-total 命令後，所有資料都會有序彙總到output-total檔案中。

getmerge會按照part-00000、part-00001、part-00002、part-00003的順序依次把每個檔案輸出到output-total檔案中。程式碼如下：

 /** Copy all files in a directory to one output file (merge). */
  public static boolean copyMerge(FileSystem srcFS, Path srcDir, 
                                  FileSystem dstFS, Path dstFile, 
                                  boolean deleteSource,
                                  Configuration conf, String addString) throws IOException {
    dstFile = checkDest(srcDir.getName(), dstFS, dstFile, false);

    if (!srcFS.getFileStatus(srcDir).isDir())
      return false;
   
    OutputStream out = dstFS.create(dstFile);
    
    try {
      FileStatus contents[] = srcFS.listStatus(srcDir);
      for (int i = 0; i < contents.length; i++) {
        if (!contents[i].isDir()) {
          InputStream in = srcFS.open(contents[i].getPath());
          try {
            IOUtils.copyBytes(in, out, conf, false);
            if (addString!=null)
              out.write(addString.getBytes("UTF-8"));
                
          } finally {
            in.close();
          } 
        }
      }
    } finally {
      out.close();
    }

part-00000、part-00001、part-00002、part-00003的訪問順序是由namenode獲取的，對應其inode節點。

修改後的完整程式碼如下：

1. TeraInputFormat.java

/**
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

package com.undirected.graph.sort;

import java.io.IOException;
import java.util.ArrayList;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.NullWritable;
import org.apache.hadoop.io.SequenceFile;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.FileInputFormat;
import org.apache.hadoop.mapred.FileSplit;
import org.apache.hadoop.mapred.InputSplit;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.LineRecordReader;
import org.apache.hadoop.mapred.RecordReader;
import org.apache.hadoop.mapred.Reporter;
import org.apache.hadoop.util.IndexedSortable;
import org.apache.hadoop.util.QuickSort;

/**
 * An input format that reads the first 10 characters of each line as the key
 * and the rest of the line as the value. Both key and value are represented
 * as Text.
 */
public class TeraInputFormat extends FileInputFormat<LongWritable,Text> {

  static final String PARTITION_FILENAME = "_partition.lst";
  static final String SAMPLE_SIZE = "terasort.partitions.sample";
  private static JobConf lastConf = null;
  private static InputSplit[] lastResult = null;

  static class TextSampler implements IndexedSortable {
    private ArrayList<LongWritable> records = new ArrayList<LongWritable>();

    public int compare(int i, int j) {
      LongWritable left = records.get(i);
      LongWritable right = records.get(j);
      return left.compareTo(right);
    }

    public void swap(int i, int j) {
      LongWritable left = records.get(i);
      LongWritable right = records.get(j);
      records.set(j, left);
      records.set(i, right);
    }

    public void addKey(LongWritable key) {
      records.add(key);
    }

    /**
     * Find the split points for a given sample. The sample keys are sorted
     * and down sampled to find even split points for the partitions. The
     * returned keys should be the start of their respective partitions.
     * @param numPartitions the desired number of partitions
     * @return an array of size numPartitions - 1 that holds the split points
     */
    LongWritable[] createPartitions(int numPartitions) {
      int numRecords = records.size();
      System.out.println("Making " + numPartitions + " from " + numRecords + 
                         " records");
      if (numPartitions > numRecords) {
        throw new IllegalArgumentException
          ("Requested more partitions than input keys (" + numPartitions +
           " > " + numRecords + ")");
      }
      
      new QuickSort().sort(this, 0, records.size());
      float stepSize = numRecords / (float) numPartitions;
      System.out.println("Step size is " + stepSize);
      LongWritable[] result = new LongWritable[numPartitions-1];
      for(int i=1; i < numPartitions; ++i) {
        result[i-1] = records.get(Math.round(stepSize * i));
      }
      // System.out.println("result :"+Arrays.toString(result));
      return result;
    }
  }
  
  /**
   * Use the input splits to take samples of the input and generate sample
   * keys. By default reads 100,000 keys from 10 locations in the input, sorts
   * them and picks N-1 keys to generate N equally sized partitions.
   * @param conf the job to sample
   * @param partFile where to write the output file to
   * @throws IOException if something goes wrong
   */
  public static void writePartitionFile(JobConf conf, 
                                        Path partFile) throws IOException {
    TeraInputFormat inFormat = new TeraInputFormat();
    TextSampler sampler = new TextSampler();
    LongWritable key = new LongWritable();
    Text value = new Text();
    int partitions = conf.getNumReduceTasks();
    long sampleSize = conf.getLong(SAMPLE_SIZE, 100000);
    InputSplit[] splits = inFormat.getSplits(conf, conf.getNumMapTasks());
    int samples = Math.min(10, splits.length);
    long recordsPerSample = sampleSize / samples;
    int sampleStep = splits.length / samples;
    long records = 0;
    // take N samples from different parts of the input
    for(int i=0; i < samples; ++i) {
      RecordReader<LongWritable,Text> reader = 
        inFormat.getRecordReader(splits[sampleStep * i], conf, null);
      while (reader.next(key, value)) {
        sampler.addKey(key);
        key=new LongWritable();
        records += 1;
        if ((i+1) * recordsPerSample <= records) {
          break;
        }
      }
    }
    FileSystem outFs = partFile.getFileSystem(conf);
    if (outFs.exists(partFile)) {
      outFs.delete(partFile, false);
    }
    SequenceFile.Writer writer = 
      SequenceFile.createWriter(outFs, conf, partFile, LongWritable.class, 
                                NullWritable.class);
    NullWritable nullValue = NullWritable.get();
    for(LongWritable split : sampler.createPartitions(partitions)) {
      writer.append(split, nullValue);
    }
    writer.close();
  }

  static class TeraRecordReader implements RecordReader<LongWritable,Text> {
    private LineRecordReader in;
    private LongWritable junk = new LongWritable();
    private Text line = new Text();

    public TeraRecordReader(Configuration job, 
                            FileSplit split) throws IOException {
      in = new LineRecordReader(job, split);
    }

    public void close() throws IOException {
      in.close();
    }

    public LongWritable createKey() {
      return new LongWritable();
    }

    public Text createValue() {
      return new Text();
    }

    public long getPos() throws IOException {
      return in.getPos();
    }

    public float getProgress() throws IOException {
      return in.getProgress();
    }

    public boolean next(LongWritable key, Text value) throws IOException {
      if (in.next(junk, line)) {
        String[] temp=line.toString().split("\t");
        key.set(Long.parseLong(temp[0]));
        if(temp.length!=1) {
          value.set(line.toString().substring(temp[0].length()+1));
        } else {
          value.set("");
        }
        return true;
      } else {
        return false;
      }
    }
  }

  @Override
  public RecordReader<LongWritable, Text> 
      getRecordReader(InputSplit split,
                      JobConf job, 
                      Reporter reporter) throws IOException {
    return new TeraRecordReader(job, (FileSplit) split);
  }

  @Override
  public InputSplit[] getSplits(JobConf conf, int splits) throws IOException {
    if (conf == lastConf) {
      return lastResult;
    }
    lastConf = conf;
    lastResult = super.getSplits(conf, splits);
    return lastResult;
  }
}

2. TeraSort.java

package com.undirected.graph.sort;

import java.io.IOException;
import java.net.URI;
import java.util.ArrayList;
import java.util.List;

import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.conf.Configured;
import org.apache.hadoop.filecache.DistributedCache;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.NullWritable;
import org.apache.hadoop.io.SequenceFile;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapred.FileOutputFormat;
import org.apache.hadoop.mapred.JobClient;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.mapred.Partitioner;
import org.apache.hadoop.mapred.TextOutputFormat;
import org.apache.hadoop.util.Tool;
import org.apache.hadoop.util.ToolRunner;


public class TeraSort extends Configured implements Tool{
  private static final Log LOG = LogFactory.getLog(TeraSort.class);
  
  /**
   * A partitioner that splits text keys into roughly equal partitions
   * in a global sorted order.
   */
  static class TotalOrderPartitioner implements Partitioner<LongWritable,Text>{

    private LongWritable[] splitPoints;
    
    /**
     * Read the cut points from the given sequence file.
     * @param fs the file system
     * @param p the path to read
     * @param job the job config
     * @return the strings to split the partitions on
     * @throws IOException
     */
    private static LongWritable[] readPartitions(FileSystem fs, Path p, 
                                         JobConf job) throws IOException {
      SequenceFile.Reader reader = new SequenceFile.Reader(fs, p, job);
      List<LongWritable> parts = new ArrayList<LongWritable>();
      LongWritable key = new LongWritable();
      NullWritable value = NullWritable.get();
      while (reader.next(key, value)) {
        parts.add(key);
        key = new LongWritable();
      }
      reader.close();
      return parts.toArray(new LongWritable[parts.size()]);  
    }
    
    @Override
    public void configure(JobConf job) {
      try {
        FileSystem fs = FileSystem.getLocal(job);
        Path partFile = new Path(TeraInputFormat.PARTITION_FILENAME);
        splitPoints = readPartitions(fs, partFile, job);
      } catch (IOException ie) {
        throw new IllegalArgumentException("can't read paritions file", ie);
      }
    }

    @Override
    public int getPartition(LongWritable key, Text value, int numPartitions) {
      if(key.get()<splitPoints[0].get()) {
        return 0;
      }
      
      for(int i=0;i<splitPoints.length-1;i++) {
        if(key.get()>=splitPoints[i].get() && key.get()<splitPoints[i+1].get()) {
           return i+1;
        }
      }
      return splitPoints.length;
    }
  }
  
  @Override
  public int run(String[] args) throws Exception {
    LOG.info("starting");
    JobConf job = (JobConf) getConf();
    Path inputDir = new Path(args[0]);
    inputDir = inputDir.makeQualified(inputDir.getFileSystem(job));
    Path partitionFile = new Path(inputDir, TeraInputFormat.PARTITION_FILENAME);
    URI partitionUri = new URI(partitionFile.toString() +
                               "#" + TeraInputFormat.PARTITION_FILENAME);
    TeraInputFormat.setInputPaths(job, new Path(args[0]));
    FileOutputFormat.setOutputPath(job, new Path(args[1]));
    job.setNumReduceTasks(Integer.parseInt(args[2]));
    job.setJobName("TeraSort");
    job.setJarByClass(TeraSort.class);
    job.setOutputKeyClass(LongWritable.class);
    job.setOutputValueClass(Text.class);
    job.setInputFormat(TeraInputFormat.class);
    job.setOutputFormat(TextOutputFormat.class);
    job.setPartitionerClass(TotalOrderPartitioner.class);
    TeraInputFormat.writePartitionFile(job, partitionFile);
    DistributedCache.addCacheFile(partitionUri, job);
    DistributedCache.createSymlink(job);
    job.setInt("dfs.replication", 1);
    JobClient.runJob(job);
    LOG.info("done");
    return 0;
  }
  /**
   * @param args
   */
  public static void main(String[] args) throws Exception {
    if(args.length<3) {
      System.out.println("Usage:<input> <output> <reduce number>");
      System.exit(-1);
    }
    int res = ToolRunner.run(new JobConf(), new TeraSort(), args);
    System.exit(res);
  }
}