上篇我刚刚学习完,Spilt的过程,还算比较简单的了,接下来学习的就是Map操作的过程了,Map和Reduce一样,是整个MapReduce的重要内容,所以,这一篇,我会好好的讲讲里面的内部实现过程。首先要说,MapTask,分为4种,可能这一点上有人就可能知道了,分别是Job-setup Task,Job-cleanup Task,Task-cleanup和Map Task。前面3个都是辅助性质的任务,不是本文分析的重点,我讲的就是里面的最最重要的MapTask。

MapTask的整个过程分为5个阶段:

Read----->Map------>Collect------->Spill------>Combine

来张时序图,简单明了:

Map Task内部实现分析_数据

在后面的代码分析中,你会看到各自方法的调用过程。

在分析整个过程之前,得先了解里面的一些内部结构,MapTask类作为Map Task的一个载体,他的类关系如下:

Map Task内部实现分析_环形缓冲_02

我们调用的就是里面的run方法,开启map任务,相应的代码:


/**
* mapTask主要执行流程
*/
@Override
public void run(final JobConf job, final TaskUmbilicalProtocol umbilical)
throws IOException, ClassNotFoundException, InterruptedException {
this.umbilical = umbilical;

// start thread that will handle communication with parent
//发送task任务报告,与父进程做交流
TaskReporter reporter = new TaskReporter(getProgress(), umbilical,
jvmContext);
reporter.startCommunicationThread();
//判断用的是新的MapReduceAPI还是旧的API
boolean useNewApi = job.getUseNewMapper();
initialize(job, getJobID(), reporter, useNewApi);

// check if it is a cleanupJobTask
//map任务有4种,Job-setup Task, Job-cleanup Task, Task-cleanup Task和MapTask
if (jobCleanup) {
//这里执行的是Job-cleanup Task
runJobCleanupTask(umbilical, reporter);
return;
}
if (jobSetup) {
//这里执行的是Job-setup Task
runJobSetupTask(umbilical, reporter);
return;
}
if (taskCleanup) {
//这里执行的是Task-cleanup Task
runTaskCleanupTask(umbilical, reporter);
return;
}

//如果前面3个任务都不是,执行的就是最主要的MapTask,根据新老API调用不同的方法
if (useNewApi) {
runNewMapper(job, splitMetaInfo, umbilical, reporter);
} else {
//我们关注一下老的方法实现splitMetaInfo为Spilt分片的信息,由于上步骤的InputFormat过程传入的
runOldMapper(job, splitMetaInfo, umbilical, reporter);
}
done(umbilical, reporter);
}

在这里我研究的都是旧的API所以往runOldMapper里面跳。在这里我要插入一句,后面的执行都会围绕着一个叫Mapper的东西,就是用户执行map函数的一个代理称呼一样,他可以完全自己重写map的背后的过程,也可以用系统自带的mapp流程。

Map Task内部实现分析_mapreduce_03

系统已经给了MapRunner的具体实现:


public void run(RecordReader<K1, V1> input, OutputCollector<K2, V2> output,
Reporter reporter)
throws IOException {
try {
// allocate key & value instances that are re-used for all entries
K1 key = input.createKey();
V1 value = input.createValue();

//从RecordReader中获取每个键值对,调用用户写的map函数
while (input.next(key, value)) {
// map pair to output
//调用用户写的map函数
mapper.map(key, value, output, reporter);
if(incrProcCount) {
reporter.incrCounter(SkipBadRecords.COUNTER_GROUP,
SkipBadRecords.COUNTER_MAP_PROCESSED_RECORDS, 1);
}
}
} finally {
//结束了关闭mapper
mapper.close();
}
}

从这里我们可以看出Map的过程就是迭代式的重复的执行用户定义的Map函数操作。好了,有了这些前提,我们可以往里深入的学习了刚刚说到了runOldMapper方法,里面马上要进行的就是Map Task的第一个过程Read。

Read阶段的作业就是从RecordReader中读取出一个个key-value,准备给后面的map过程执行map函数操作。


//获取输入inputSplit信息
InputSplit inputSplit = getSplitDetails(new Path(splitIndex.getSplitLocation()),
splitIndex.getStartOffset());

updateJobWithSplit(job, inputSplit);
reporter.setInputSplit(inputSplit);

//是否是跳过错误记录模式,获取RecordReader
RecordReader<INKEY,INVALUE> in = isSkipping() ?
new SkippingRecordReader<INKEY,INVALUE>(inputSplit, umbilical, reporter) :
new TrackedRecordReader<INKEY,INVALUE>(inputSplit, job, reporter);

后面的就是Map阶段,把值取出来之后,就要给Mapper去执行里面的run方法了,run方法里面会调用用户自己实现的map函数,之前也都是分析过了的。在用户编写的map的尾部,一般会调用collect.collect()方法,把处理后的key-value输出,这个时候,也就来到了collect阶段。


runner.run(in, new OldOutputCollector(collector, conf), reporter);

之后进行的是Collect阶段主要的操作时什么呢,就是把一堆堆的key-value进行分区输出到环形缓冲区中,这是的数据仅仅放在内存中,还没有写到磁盘中。在collect这个过程中涉及的东西还比较多,看一下结构关系图;

Map Task内部实现分析_hdfs_04


里面有个partitioner的成员变量,专门用于获取key-value的的分区号,默认是通过key的哈希取模运算,得到分区号的,当然你可以自定义实现,如果不分区的话partition就是等于-1。


/**
* Since the mapred and mapreduce Partitioners don't share a common interface
* (JobConfigurable is deprecated and a subtype of mapred.Partitioner), the
* partitioner lives in Old/NewOutputCollector. Note that, for map-only jobs,
* the configured partitioner should not be called. It's common for
* partitioners to compute a result mod numReduces, which causes a div0 error
*/
private static class OldOutputCollector<K,V> implements OutputCollector<K,V> {
private final Partitioner<K,V> partitioner;
private final MapOutputCollector<K,V> collector;
private final int numPartitions;

@SuppressWarnings("unchecked")
OldOutputCollector(MapOutputCollector<K,V> collector, JobConf conf) {
numPartitions = conf.getNumReduceTasks();
if (numPartitions > 0) {
//如果分区数大于0,则反射获取系统配置方法,默认哈希去模,用户可以自己实现字节的分区方法
//因为是RPC传来的,所以采用反射
partitioner = (Partitioner<K,V>)
ReflectionUtils.newInstance(conf.getPartitionerClass(), conf);
} else {
//如果分区数为0,说明不进行分区
partitioner = new Partitioner<K,V>() {
@Override
public void configure(JobConf job) { }
@Override
public int getPartition(K key, V value, int numPartitions) {
//分区号直接返回-1代表不分区处理
return -1;
}
};
}
this.collector = collector;
}
.....

collect的代理调用实现方法如下,注意此时还不是真正调用:


.....
@Override
public void collect(K key, V value) throws IOException {
try {
//具体通过collect方法分区写入内存,调用partitioner.getPartition获取分区号
//缓冲区为环形缓冲区
collector.collect(key, value,
partitioner.getPartition(key, value, numPartitions));
} catch (InterruptedException ie) {
Thread.currentThread().interrupt();
throw new IOException("interrupt exception", ie);
}
}

这里的collector指的是上面代码中的MapOutputCollector对象,开放给用调用的是OldOutputCollector,但是我们看看代码:


interface MapOutputCollector<K, V> {

public void collect(K key, V value, int partition
) throws IOException, InterruptedException;
public void close() throws IOException, InterruptedException;

public void flush() throws IOException, InterruptedException,
ClassNotFoundException;

}


他只是一个接口,真正的实现是谁呢?这个时候应该回头看一下代码:


private <INKEY,INVALUE,OUTKEY,OUTVALUE>
void runOldMapper(final JobConf job,
final TaskSplitIndex splitIndex,
final TaskUmbilicalProtocol umbilical,
TaskReporter reporter
) throws IOException, InterruptedException,
ClassNotFoundException {
...
int numReduceTasks = conf.getNumReduceTasks();
LOG.info("numReduceTasks: " + numReduceTasks);
MapOutputCollector collector = null;
if (numReduceTasks > 0) {
//如果存在ReduceTask,则将数据存入MapOutputBuffer环形缓冲
collector = new MapOutputBuffer(umbilical, job, reporter);
} else {
//如果没有ReduceTask任务的存在,直接写入把操作结果写入HDFS作为最终结果
collector = new DirectMapOutputCollector(umbilical, job, reporter);
}
MapRunnable<INKEY,INVALUE,OUTKEY,OUTVALUE> runner =
ReflectionUtils.newInstance(job.getMapRunnerClass(), job);

try {
runner.run(in, new OldOutputCollector(collector, conf), reporter);
.....

分为2种情况当有Reduce任务时,collector为MapOutputBuffer,没有Reduce任务时为DirectMapOutputCollector,从这里也能明白,作者考虑的很周全呢,没有Reduce直接写入HDFS,效率会高很多。也就是说,最终的collect方法就是MapOutputBuffer的方法了。

因为collect的操作时将数据存入环形缓冲区,这意味着,用户对数据的读写都是在同个缓冲区上的,所以为了避免出现脏数据的现象,一定会做额外处理,这里作者用了和BlockingQueue类似的操作,用一个ReetrantLocj,获取2个锁控制条件,一个为spillDone

,一个为spillReady,同个condition的await,signal方法实现丢缓冲区的读写控制。


.....
private final ReentrantLock spillLock = new ReentrantLock();
private final Condition spillDone = spillLock.newCondition();
private final Condition spillReady = spillLock.newCondition();
.....

然后看collect的方法:


public synchronized void collect(K key, V value, int partition
) throws IOException {
.....
try {
// serialize key bytes into buffer
int keystart = bufindex;
keySerializer.serialize(key);
if (bufindex < keystart) {
// wrapped the key; reset required
bb.reset();
keystart = 0;
}
// serialize value bytes into buffer
final int valstart = bufindex;
valSerializer.serialize(value);
int valend = bb.markRecord();

if (partition < 0 || partition >= partitions) {
throw new IOException("Illegal partition for " + key + " (" +
partition + ")");
}
....


至于环形缓冲区的结构,不是本文的重点,结构设计还是比较复杂的,大家可以自行学习。当环形缓冲区内的数据渐渐地被填满之后,会出现"溢写"操作,就是把缓冲中的数据写到磁盘DISK中,这个过程就是后面的Spill阶段了。

Spill的阶段会时不时的穿插在collect的执行过程中。


...
if (kvstart == kvend && kvsoftlimit) {
LOG.info("Spilling map output: record full = " + kvsoftlimit);
startSpill();
}

如果开头kvstart的位置等kvend的位置,说明转了一圈有到头了,数据已经满了的状态,开始spill溢写操作。


private synchronized void startSpill() {
LOG.info("bufstart = " + bufstart + "; bufend = " + bufmark +
"; bufvoid = " + bufvoid);
LOG.info("kvstart = " + kvstart + "; kvend = " + kvindex +
"; length = " + kvoffsets.length);
kvend = kvindex;
bufend = bufmark;
spillReady.signal();
}

会触发condition的信号量操作:


private synchronized void startSpill() {
LOG.info("bufstart = " + bufstart + "; bufend = " + bufmark +
"; bufvoid = " + bufvoid);
LOG.info("kvstart = " + kvstart + "; kvend = " + kvindex +
"; length = " + kvoffsets.length);
kvend = kvindex;
bufend = bufmark;
spillReady.signal();
}

就会跑到了SpillThead这个地方执行sortAndSpill方法:


spillThreadRunning = true;
try {
while (true) {
spillDone.signal();
while (kvstart == kvend) {
spillReady.await();
}
try {
spillLock.unlock();
//当缓冲区溢出时,写到磁盘中
sortAndSpill();

sortAndSpill里面会对数据做写入文件操作写入之前还会有sort排序操作,数据多了还会进行一定的combine合并操作。


private void sortAndSpill() throws IOException, ClassNotFoundException,
InterruptedException {
......
try {
// create spill file
final SpillRecord spillRec = new SpillRecord(partitions);
final Path filename =
mapOutputFile.getSpillFileForWrite(numSpills, size);
out = rfs.create(filename);

final int endPosition = (kvend > kvstart)
? kvend
: kvoffsets.length + kvend;
//在写入操作前进行排序操作
sorter.sort(MapOutputBuffer.this, kvstart, endPosition, reporter);
int spindex = kvstart;
IndexRecord rec = new IndexRecord();
InMemValBytes value = new InMemValBytes();
for (int i = 0; i < partitions; ++i) {
IFile.Writer<K, V> writer = null;
try {
long segmentStart = out.getPos();
writer = new Writer<K, V>(job, out, keyClass, valClass, codec,
spilledRecordsCounter);
if (combinerRunner == null) {
// spill directly
DataInputBuffer key = new DataInputBuffer();
while (spindex < endPosition &&
kvindices[kvoffsets[spindex % kvoffsets.length]
+ PARTITION] == i) {
final int kvoff = kvoffsets[spindex % kvoffsets.length];
getVBytesForOffset(kvoff, value);
key.reset(kvbuffer, kvindices[kvoff + KEYSTART],
(kvindices[kvoff + VALSTART] -
kvindices[kvoff + KEYSTART]));
//writer中写入键值对操作
writer.append(key, value);
++spindex;
}
} else {
int spstart = spindex;
while (spindex < endPosition &&
kvindices[kvoffsets[spindex % kvoffsets.length]
+ PARTITION] == i) {
++spindex;
}
// Note: we would like to avoid the combiner if we've fewer
// than some threshold of records for a partition
//如果分区多的话,执行合并操作
if (spstart != spindex) {
combineCollector.setWriter(writer);
RawKeyValueIterator kvIter =
new MRResultIterator(spstart, spindex);
//执行一次文件合并combine操作
combinerRunner.combine(kvIter, combineCollector);
}
}

......
//写入到文件中
spillRec.writeToFile(indexFilename, job);
} else {
indexCacheList.add(spillRec);
totalIndexCacheMemory +=
spillRec.size() * MAP_OUTPUT_INDEX_RECORD_LENGTH;
}
LOG.info("Finished spill " + numSpills);
++numSpills;
} finally {
if (out != null) out.close();
}
}

每次Spill的过程都会产生一堆堆的文件,在最后的时候就会来到了Combine阶段,也就是Map任务的最后一个阶段了,他的任务就是把所有上一阶段的任务产生的文件进行Merge操作,合并成一个文件,便于后面的Reduce的任务的读取,在代码的对应实现中是collect.flush()方法。


.....
try {
runner.run(in, new OldOutputCollector(collector, conf), reporter);
//将collector中的数据刷新到内存中去
collector.flush();
} finally {
//close
in.close(); // close input
collector.close();
}
}

这里的collector的flush方法调用的就是MapOutputBuffer.flush方法,

public synchronized void flush() throws IOException, ClassNotFoundException,
InterruptedException {
...
// shut down spill thread and wait for it to exit. Since the preceding
// ensures that it is finished with its work (and sortAndSpill did not
// throw), we elect to use an interrupt instead of setting a flag.
// Spilling simultaneously from this thread while the spill thread
// finishes its work might be both a useful way to extend this and also
// sufficient motivation for the latter approach.
try {
spillThread.interrupt();
spillThread.join();
} catch (InterruptedException e) {
throw (IOException)new IOException("Spill failed"
).initCause(e);
}
// release sort buffer before the merge
kvbuffer = null;
//最后进行merge合并成一个文件
mergeParts();
Path outputPath = mapOutputFile.getOutputFile();
fileOutputByteCounter.increment(rfs.getFileStatus(outputPath).getLen());
}

至此,Map任务宣告结束了,整体流程还是真是有点九曲十八弯的感觉。分析这么一个比较庞杂的过程,我一直在想如何更好的表达出我的想法,