雪山飞狐

浏览: 55099 次
性别:
来自: 广西

最近访客更多访客>>

co1de

zhangwanhua

youcp999

RoamsWind

博主相关

博客

微博

相册

留言

关于我

文章分类

社区版块

存档分类

Mapper输入InputSplit分片数透析

博客分类：

Hadoop

什么是InputSplit

InputSplit是指分片，在MapReduce当中作业中，作为map task最小输入单位。分片是基于文件基础上出来的而来的概念，通俗的理解一个文件可以切分为多少个片段，每个片段包括了<文件名，开始位置，长度，位于哪些主机>等信息。map task的数量由输入文件总大小和分片大小确定的；hadoop2.2版本hdfs的数据块默认是128M。若一个文件大于128M，通过将大文件分解得到若干个数据块；若一个文件小于128M，则按它的实际大小组块存储；

因此可以解释为什么hdfs喜欢海量的大文件，而不喜欢海量的小文件：

（1）海量的小文件需要namenode存储较多的元数据信息记录块的位置（假如一个64M的数据块（大文件分解的块）需要1k的元数据信息，那么1M的数据块（小文件的块）仍需要1k的元数据信息）

（2）HDFS就是为流式访问大文件而设计的，如果访问大量小文件，需要不断的从一个datanode跳到另一个datanode，严重影响性能。

hdfs的一个数据块一定属于一个文件；

这里又有个问题：为什么map task分片大小最好和hdfs设置的块大小一致呢？

我们综合考虑两点：

（1）map task的个数=输入文件总大小（totalSize）/分片尺寸(splitSize)。也就是说分片尺寸越大，map task的个数就越少=>系统执行的开销越小，系统管理分片的开销越小。

（2）网络传输开销，如果分片太大以至于一个分片要跨越多个HDFS块，则一个map任务必须要由多个块通过网络传输，所以分片大小的上限是HDFS块的大小。

所以，map任务时的分片大小设置为HDFS块的大小是最佳选择。

流程

FileInputFormat.getSplits返回文件的分片数目，这部分将介绍其运行流程，后面将粘贴其源码并给出注释

通过listStatus()获取输入文件列表files，其中会遍历输入目录的子目录，并过滤掉部分文件，如文件_SUCCESS
获取所有的文件大小totalSIze
goalSIze=totalSize/numMaps。numMaps是用户指定的map数目
files中取出一个文件file
计算splitSize。splitSize=max(minSplitSize,min(file.blockSize,goalSize))，其中minSplitSize是允许的最小分片大小，默认为1B
后面根据splitSize大小将file分片。在分片的时候，如果剩余的大小不大于splitSize*1.1，且大于0B的时候，会将该区域整个作为一个分片。这样做是为了防止一个mapper处理的数据太小
将file的分片加入到splits中
返回4，直到将files遍历完
结束，返回splits

源码(hadoop2.2.0)

其实流程算起来也不算复杂，所以就直接用代码注释来做吧

这里边涉及这么几个方法：

1、public List<InputSplit> getSplits(JobContext job), 这个由客户端调用来获得当前Job的所有分片(split)，然后发送给JobTracker（新API中应该是ResourceManager），而JobTracker根据这些分片的存储位置来给TaskTracker分配map任务去处理这些分片。这个方法用到了后边的listStatus，然后根据得到的这些文件信息，从FileSystem那里去拉取这些组成这些文件的块的信息(BlockLocation)，使用的是getFileBlockLocation(file,start,len)，这个方法是与使用的文件系统实现相关的(FileSystem，LocalFileSystem，DistributedFileSystem)

	/** 
	   * Generate the list of files and make them into FileSplits.
	   * @param job the job context
	   * @throws IOException
	   */
	  public List<InputSplit> getSplits(JobContext job) throws IOException {
		//根据默认配置与job所配置minInputSplitSize取得最小切片大小
	    long minSize = Math.max(getFormatMinSplitSize(), getMinSplitSize(job));
	    //根据job所配置maxInputSplitSize取得最大切片大小
	    long maxSize = getMaxSplitSize(job);
	 
	    //保存生成的split信息
	    List<InputSplit> splits = new ArrayList<InputSplit>();
	    List<FileStatus> files = listStatus(job);//获取job所配置的文件列表
	    //对每个文件进行切片
	    for (FileStatus file: files) {
	      Path path = file.getPath();
	      long length = file.getLen();
	      if (length != 0) {
	        BlockLocation[] blkLocations;
	        if (file instanceof LocatedFileStatus) {
	          blkLocations = ((LocatedFileStatus) file).getBlockLocations();
	        } else {
	          FileSystem fs = path.getFileSystem(job.getConfiguration());
	          blkLocations = fs.getFileBlockLocations(file, 0, length);
	        }
	        if (isSplitable(job, path)) {//验证非二进制文件可切分（二进制文件不允许切分）
	          long blockSize = file.getBlockSize();
	          long splitSize = computeSplitSize(blockSize, minSize, maxSize);//计算分块尺寸
	 
	          long bytesRemaining = length;
	          //循环按分片大小取出一个个分片 
	          while (((double) bytesRemaining)/splitSize > SPLIT_SLOP) {
	            int blkIndex = getBlockIndex(blkLocations, length-bytesRemaining);//计算分块索引
	            splits.add(makeSplit(path, length-bytesRemaining, splitSize,
	                                     blkLocations[blkIndex].getHosts()));
	            bytesRemaining -= splitSize;
	          }
	          //对尾部不足一个分片大小的也生成一个分片
	          if (bytesRemaining != 0) {
	            int blkIndex = getBlockIndex(blkLocations, length-bytesRemaining);
	            splits.add(makeSplit(path, length-bytesRemaining, bytesRemaining,
	                       blkLocations[blkIndex].getHosts()));
	          }
	        } else { // not splitable
	          //不可分块文件整体作为一块
	          splits.add(makeSplit(path, 0, length, blkLocations[0].getHosts()));
	        }
	      } else { //空文件占用一个分块
	        //Create empty hosts array for zero length files
	        splits.add(makeSplit(path, 0, length, new String[0]));
	      }
	    }
	    // Save the number of input files for metrics/loadgen
	    job.getConfiguration().setLong(NUM_INPUT_FILES, files.size());
	    LOG.debug("Total # of splits: " + splits.size());
	    return splits;
	  }

2、protected List<FileStatus> listStatus(JobContext job), 先根据“mapred.input.dir”的配置值去得到用户指定的所有Path。然后根据这个JobContext的Configuration得到FileSystem(当然，更可能是 DistributedFileSystem )。最后应用用户可能设置了的PathFilter，通过FileSystem获取所有这些Path所代表的File(FileStatus)。注：这个方法的东西相当多，很多内容还十分陌生。

/** List input directories.
   * Subclasses may override to, e.g., select only files matching a regular
   * expression. 
   * 
   * @param job the job to list input paths for
   * @return array of FileStatus objects
   * @throws IOException if zero items.
   */
  protected List<FileStatus> listStatus(JobContext job
                                        ) throws IOException {
    List<FileStatus> result = new ArrayList<FileStatus>();
    Path[] dirs = getInputPaths(job);
    if (dirs.length == 0) {
      throw new IOException("No input paths specified in job");
    }
     
    // get tokens for all the required FileSystems..
    TokenCache.obtainTokensForNamenodes(job.getCredentials(), dirs, 
                                        job.getConfiguration());
 
    // Whether we need to recursive look into the directory structure
    boolean recursive = getInputDirRecursive(job);
     
    List<IOException> errors = new ArrayList<IOException>();
     
    // creates a MultiPathFilter with the hiddenFileFilter and the
    // user provided one (if any).
    List<PathFilter> filters = new ArrayList<PathFilter>();
    filters.add(hiddenFileFilter);
    PathFilter jobFilter = getInputPathFilter(job);
    if (jobFilter != null) {
      filters.add(jobFilter);
    }
    PathFilter inputFilter = new MultiPathFilter(filters);
     
    for (int i=0; i < dirs.length; ++i) {
      Path p = dirs[i];
      FileSystem fs = p.getFileSystem(job.getConfiguration()); 
      FileStatus[] matches = fs.globStatus(p, inputFilter);
      if (matches == null) {
        errors.add(new IOException("Input path does not exist: " + p));
      } else if (matches.length == 0) {
        errors.add(new IOException("Input Pattern " + p + " matches 0 files"));
      } else {
        for (FileStatus globStat: matches) {
          if (globStat.isDirectory()) {
            RemoteIterator<LocatedFileStatus> iter =
                fs.listLocatedStatus(globStat.getPath());
            while (iter.hasNext()) {
              LocatedFileStatus stat = iter.next();
              if (inputFilter.accept(stat.getPath())) {
                if (recursive && stat.isDirectory()) {
                  addInputPathRecursively(result, fs, stat.getPath(),
                      inputFilter);
                } else {
                  result.add(stat);
                }
              }
            }
          } else {
            result.add(globStat);
          }
        }
      }
    }
 
    if (!errors.isEmpty()) {
      throw new InvalidInputException(errors);
    }
    LOG.info("Total input paths to process : " + result.size()); 
    return result;
  }

3、protected long computeSplitSize(long blockSize, long minSize, long maxSize),计算出当前Job所配置的分片最大尺寸。

protected long computeSplitSize(long blockSize, long minSize,
                                  long maxSize) {
    return Math.max(minSize, Math.min(maxSize, blockSize));
  }

4、protected int getBlockIndex(BlockLocation[] blkLocations, long offset), 由于组成文件的块的信息已经获得了，只需要根据offset来计算所在的那个块就行了。

  protected int getBlockIndex(BlockLocation[] blkLocations, 
                              long offset) {
    for (int i = 0 ; i < blkLocations.length; i++) {
      // is the offset inside this block?
      if ((blkLocations[i].getOffset() <= offset) &&
          (offset < blkLocations[i].getOffset() + blkLocations[i].getLength())){
        return i;
      }
    }
    BlockLocation last = blkLocations[blkLocations.length -1];
    long fileLength = last.getOffset() + last.getLength() -1;
    throw new IllegalArgumentException("Offset " + offset + 
                                       " is outside of file (0.." +
                                       fileLength + ")");
  }

5、通过源码的分析我们知道可以通过Configuration conf设置 “mapred.max.split.size”和“mapred.min.split.size”或者Job上设置maxInputSplitSize和minInputSplitSize的值进行调整InputSplit切片的大小最终达到调整map task数量的目的。

实例

下面通过一个简单实例测试

三个文件mapperstest1-01.txt，mapperstest01-02.txt，mapperstest1-03.txt

1、没配置相关minSplitSize和maxSplitSize，即使用默认配置，splitSize=blockSize=128M。

代码

package com.ljf.mr.test;

import java.io.IOException;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapreduce.Job;
import org.apache.hadoop.mapreduce.Mapper;
import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
import org.apache.hadoop.util.GenericOptionsParser;

public class MappersTest {

	public static class Mappers extends Mapper<LongWritable,Text,IntWritable,Text> {
		
		private static IntWritable intKey = new IntWritable(1);
		
		/**
		 * Text文本缺省使用LineRecordReader类读取
		 * 一行一个key/value对，key取文本偏移量
		 * value为行的文本内容
		 * @throws InterruptedException 
		 * @throws IOException 
		 */
		public void map(LongWritable key, Text value, Context context) throws IOException, InterruptedException {
		    context.write(intKey, value);
		}
	}
	
	public static void main(String[] args) throws IOException, ClassNotFoundException, InterruptedException {
		if(args.length < 2) {
			System.exit(2);
		}
		Configuration conf = new Configuration();
		
		String[] otherArgs = new GenericOptionsParser(conf, args).getRemainingArgs();

		@SuppressWarnings("deprecation")
		Job job = new Job(conf, "MappersTest");
		
		job.setJarByClass(MappersTest.class);
		job.setMapperClass(Mappers.class);
		
		job.setMapOutputKeyClass(IntWritable.class);
		job.setMapOutputValueClass(Text.class);
		
		FileInputFormat.addInputPath(job, new Path(otherArgs[0]));
		FileOutputFormat.setOutputPath(job, new Path(otherArgs[1]));
		
		System.exit(job.waitForCompletion(true)? 0:1);
	}

}

运行过程

14/09/25 19:21:22 INFO input.FileInputFormat: Total input paths to process : 3
14/09/25 19:21:22 INFO mapreduce.JobSubmitter: number of splits:3
14/09/25 19:21:22 INFO Configuration.deprecation: user.name is deprecated. Instead, use mapreduce.job.user.name
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.jar is deprecated. Instead, use mapreduce.job.jar
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.mapoutput.value.class is deprecated. Instead, use mapreduce.map.output.value.class
14/09/25 19:21:22 INFO Configuration.deprecation: mapreduce.map.class is deprecated. Instead, use mapreduce.job.map.class
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.job.name is deprecated. Instead, use mapreduce.job.name
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.input.dir is deprecated. Instead, use mapreduce.input.fileinputformat.inputdir
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.output.dir is deprecated. Instead, use mapreduce.output.fileoutputformat.outputdir
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.map.tasks is deprecated. Instead, use mapreduce.job.maps
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.mapoutput.key.class is deprecated. Instead, use mapreduce.map.output.key.class
14/09/25 19:21:22 INFO Configuration.deprecation: mapred.working.dir is deprecated. Instead, use mapreduce.job.working.dir
14/09/25 19:21:23 INFO mapreduce.JobSubmitter: Submitting tokens for job: job_1411613494956_0004
14/09/25 19:21:23 INFO impl.YarnClientImpl: Submitted application application_1411613494956_0004 to ResourceManager at /0.0.0.0:8032
14/09/25 19:21:23 INFO mapreduce.Job: The url to track the job: http://ubuntu:8088/proxy/application_1411613494956_0004/
14/09/25 19:21:23 INFO mapreduce.Job: Running job: job_1411613494956_0004
14/09/25 19:21:50 INFO mapreduce.Job: Job job_1411613494956_0004 running in uber mode : false
14/09/25 19:21:50 INFO mapreduce.Job:  map 0% reduce 0%
14/09/25 19:22:29 INFO mapreduce.Job:  map 100% reduce 0%
14/09/25 19:22:29 INFO mapreduce.Job: Task Id : attempt_1411613494956_0004_m_000001_0, Status : FAILED
14/09/25 19:23:32 INFO mapreduce.Job:  map 100% reduce 0%
14/09/25 19:23:33 INFO mapreduce.Job:  map 100% reduce 100%
14/09/25 19:23:34 INFO mapreduce.Job: Job job_1411613494956_0004 completed successfully
14/09/25 19:23:34 INFO mapreduce.Job: Counters: 46
	File System Counters
		FILE: Number of bytes read=19412
		FILE: Number of bytes written=354065
		FILE: Number of read operations=0
		FILE: Number of large read operations=0
		FILE: Number of write operations=0
		HDFS: Number of bytes read=12524
		HDFS: Number of bytes written=14550
		HDFS: Number of read operations=12
		HDFS: Number of large read operations=0
		HDFS: Number of write operations=2
	Job Counters 
		Failed map tasks=4
		Killed map tasks=1
		Launched map tasks=8
		Launched reduce tasks=1
		Other local map tasks=5
		Data-local map tasks=3
		Total time spent by all maps in occupied slots (ms)=246514
		Total time spent by all reduces in occupied slots (ms)=23106
	Map-Reduce Framework
		Map input records=1214
		Map output records=1214
		Map output bytes=16978
		Map output materialized bytes=19424
		Input split bytes=402
		Combine input records=0
		Combine output records=0
		Reduce input groups=1
		Reduce shuffle bytes=19424
		Reduce input records=1214
		Reduce output records=1214
		Spilled Records=2428
		Shuffled Maps =3
		Failed Shuffles=0
		Merged Map outputs=3
		GC time elapsed (ms)=925
		CPU time spent (ms)=2850
		Physical memory (bytes) snapshot=621813760
		Virtual memory (bytes) snapshot=2639740928
		Total committed heap usage (bytes)=377892864
	Shuffle Errors
		BAD_ID=0
		CONNECTION=0
		IO_ERROR=0
		WRONG_LENGTH=0
		WRONG_MAP=0
		WRONG_REDUCE=0
	File Input Format Counters 
		Bytes Read=12122
	File Output Format Counters 
		Bytes Written=14550

2、配置相关minSplitSize和maxSplitSize的值，maxSplitSize=10240(10KB)。

代码

package com.ljf.mr.test;

import java.io.IOException;

import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.io.IntWritable;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Text;
import org.apache.hadoop.mapreduce.Job;
import org.apache.hadoop.mapreduce.Mapper;
import org.apache.hadoop.mapreduce.lib.input.FileInputFormat;
import org.apache.hadoop.mapreduce.lib.output.FileOutputFormat;
import org.apache.hadoop.util.GenericOptionsParser;

public class MappersTest {

	public static class Mappers extends Mapper<LongWritable,Text,IntWritable,Text> {
		
		private static IntWritable intKey = new IntWritable(1);
		
		/**
		 * Text文本缺省使用LineRecordReader类读取
		 * 一行一个key/value对，key取文本偏移量
		 * value为行的文本内容
		 * @throws InterruptedException 
		 * @throws IOException 
		 */
		public void map(LongWritable key, Text value, Context context) throws IOException, InterruptedException {
		    context.write(intKey, value);
		}
	}
	
	public static void main(String[] args) throws IOException, ClassNotFoundException, InterruptedException {
		if(args.length < 2) {
			System.exit(2);
		}
		Configuration conf = new Configuration();
//		conf.setStrings("mapred.max.split.size", "10240");
//		conf.setStrings("mapred.min.split.size", "10240");
		
		String[] otherArgs = new GenericOptionsParser(conf, args).getRemainingArgs();

		@SuppressWarnings("deprecation")
		Job job = new Job(conf, "MappersTest");
		
		job.setJarByClass(MappersTest.class);
		job.setMapperClass(Mappers.class);
		
		job.setMapOutputKeyClass(IntWritable.class);
		job.setMapOutputValueClass(Text.class);
		
		FileInputFormat.setMaxInputSplitSize(job, 10240);//10KB
		FileInputFormat.addInputPath(job, new Path(otherArgs[0]));
		FileOutputFormat.setOutputPath(job, new Path(otherArgs[1]));
		
		System.exit(job.waitForCompletion(true)? 0:1);
	}

}

运行过程

14/09/25 19:29:53 INFO input.FileInputFormat: Total input paths to process : 3
14/09/25 19:29:53 INFO mapreduce.JobSubmitter: number of splits:4
14/09/25 19:29:53 INFO Configuration.deprecation: user.name is deprecated. Instead, use mapreduce.job.user.name
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.jar is deprecated. Instead, use mapreduce.job.jar
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.mapoutput.value.class is deprecated. Instead, use mapreduce.map.output.value.class
14/09/25 19:29:53 INFO Configuration.deprecation: mapreduce.map.class is deprecated. Instead, use mapreduce.job.map.class
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.job.name is deprecated. Instead, use mapreduce.job.name
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.input.dir is deprecated. Instead, use mapreduce.input.fileinputformat.inputdir
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.output.dir is deprecated. Instead, use mapreduce.output.fileoutputformat.outputdir
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.max.split.size is deprecated. Instead, use mapreduce.input.fileinputformat.split.maxsize
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.map.tasks is deprecated. Instead, use mapreduce.job.maps
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.mapoutput.key.class is deprecated. Instead, use mapreduce.map.output.key.class
14/09/25 19:29:53 INFO Configuration.deprecation: mapred.working.dir is deprecated. Instead, use mapreduce.job.working.dir
14/09/25 19:29:53 INFO mapreduce.JobSubmitter: Submitting tokens for job: job_1411613494956_0005
14/09/25 19:29:54 INFO impl.YarnClientImpl: Submitted application application_1411613494956_0005 to ResourceManager at /0.0.0.0:8032
14/09/25 19:29:54 INFO mapreduce.Job: The url to track the job: http://ubuntu:8088/proxy/application_1411613494956_0005/
14/09/25 19:29:54 INFO mapreduce.Job: Running job: job_1411613494956_0005
14/09/25 19:30:16 INFO mapreduce.Job: Job job_1411613494956_0005 running in uber mode : false
14/09/25 19:30:16 INFO mapreduce.Job:  map 0% reduce 0%
14/09/25 19:30:56 INFO mapreduce.Job: Task Id : attempt_1411613494956_0005_m_000000_0, Status : FAILED
Container killed on request. Exit code is 137
Killed by external signal

14/09/25 19:30:57 INFO mapreduce.Job:  map 75% reduce 0%
14/09/25 19:31:21 INFO mapreduce.Job:  map 100% reduce 0%
14/09/25 19:31:23 INFO mapreduce.Job:  map 100% reduce 100%
14/09/25 19:31:23 INFO mapreduce.Job: Job job_1411613494956_0005 completed successfully
14/09/25 19:31:23 INFO mapreduce.Job: Counters: 46
	File System Counters
		FILE: Number of bytes read=19412
		FILE: Number of bytes written=433651
		FILE: Number of read operations=0
		FILE: Number of large read operations=0
		FILE: Number of write operations=0
		HDFS: Number of bytes read=14420
		HDFS: Number of bytes written=14550
		HDFS: Number of read operations=15
		HDFS: Number of large read operations=0
		HDFS: Number of write operations=2
	Job Counters 
		Failed map tasks=1
		Killed map tasks=1
		Launched map tasks=6
		Launched reduce tasks=1
		Other local map tasks=2
		Data-local map tasks=4
		Total time spent by all maps in occupied slots (ms)=188177
		Total time spent by all reduces in occupied slots (ms)=23271
	Map-Reduce Framework
		Map input records=1214
		Map output records=1214
		Map output bytes=16978
		Map output materialized bytes=19430
		Input split bytes=536
		Combine input records=0
		Combine output records=0
		Reduce input groups=1
		Reduce shuffle bytes=19430
		Reduce input records=1214
		Reduce output records=1214
		Spilled Records=2428
		Shuffled Maps =4
		Failed Shuffles=0
		Merged Map outputs=4
		GC time elapsed (ms)=1385
		CPU time spent (ms)=4280
		Physical memory (bytes) snapshot=582656000
		Virtual memory (bytes) snapshot=3298025472
		Total committed heap usage (bytes)=498614272
	Shuffle Errors
		BAD_ID=0
		CONNECTION=0
		IO_ERROR=0
		WRONG_LENGTH=0
		WRONG_MAP=0
		WRONG_REDUCE=0
	File Input Format Counters 
		Bytes Read=13884
	File Output Format Counters 
		Bytes Written=14550

http://my.oschina.net/KingPan/blog/288549 写道

参考MapReduce Application中mapper的数目和分片的数目