MapReduce的MapTask任务的运行源码级分析 这篇文章好不容易恢复了。。。谢天谢地。。。这篇文章讲了MapTask的执行流程。咱们这一节讲解ReduceTask的执行流程。ReduceTask也有四种任务,可参考前一章节对应的内容,至于Reduce Task要从各个Map Task上读取一片数据,经过排序后,以组为单位交给用户编写的reduce方法,并将结果写入HDFS中。
MapTask和ReduceTask都是Task的子类,分别对应于我们常说的map和reduce任务。同上一节一样Child类中直接运行的是run方法,ReduceTask.run()方法代码如下:
//ReduceTask.run方法开始和MapTask类似,包括initialize()初始化,根据情况看是否调用runJobCleanupTask(),
//runJobSetupTask(),runTaskCleanupTask()。之后进入正式的工作,主要有这么三个步骤:Copy、Sort、Reduce。
@Override
@SuppressWarnings("unchecked")
public void run(JobConf job, final TaskUmbilicalProtocol umbilical)
throws IOException, InterruptedException, ClassNotFoundException {
this.umbilical = umbilical;
job.setBoolean("mapred.skip.on", isSkipping());
/*添加reduce过程需要经过的几个阶段。以便通知TaskTracker目前运 行的情况*/
if (isMapOrReduce()) {
copyPhase = getProgress().addPhase("copy");
sortPhase = getProgress().addPhase("sort");
reducePhase = getProgress().addPhase("reduce");
}
// start thread that will handle communication with parent
// 设置并启动reporter进程以便和TaskTracker进行交流
TaskReporter reporter = new TaskReporter(getProgress(), umbilical,
jvmContext);
reporter.startCommunicationThread();
//在job client中初始化job时,默认就是用新的API,详见Job.setUseNewAPI()方法
boolean useNewApi = job.getUseNewReducer();
/*用来初始化任务,主要是进行一些和任务输出相关的设置,比如创建commiter,设置工作目录等*/
initialize(job, getJobID(), reporter, useNewApi);//这里将会处理输出目录
/*以下4个if语句均是根据任务类型的不同进行相应的操作,这些方 法均是Task类的方法,所以与任务是MapTask还是ReduceTask无关*/
// check if it is a cleanupJobTask
if (jobCleanup) {
runJobCleanupTask(umbilical, reporter);
return;
}
if (jobSetup) {
//主要是创建工作目录的FileSystem对象
runJobSetupTask(umbilical, reporter);
return;
}
if (taskCleanup) {
//设置任务目前所处的阶段为结束阶段,并且删除工作目录
runTaskCleanupTask(umbilical, reporter);
return;
} // Initialize the codec
codec = initCodec(); boolean isLocal = "local".equals(job.get("mapred.job.tracker", "local")); //判断是否是单机hadoop
if (!isLocal) {
//1. Copy.就是从执行各个Map任务的服务器那里,收到map的输出文件。拷贝的任务,是由ReduceTask.ReduceCopier 类来负责。
//ReduceCopier对象负责将Map函数的输出拷贝至Reduce所在机器
reduceCopier = new ReduceCopier(umbilical, job, reporter);
if (!reduceCopier.fetchOutputs()) {////fetchOutputs函数负责拷贝各个Map函数的输出
if(reduceCopier.mergeThrowable instanceof FSError) {
throw (FSError)reduceCopier.mergeThrowable;
}
throw new IOException("Task: " + getTaskID() +
" - The reduce copier failed", reduceCopier.mergeThrowable);
}
}
copyPhase.complete(); // copy is already complete
setPhase(TaskStatus.Phase.SORT);
statusUpdate(umbilical); final FileSystem rfs = FileSystem.getLocal(job).getRaw();
//2.Sort(其实相当于合并).排序工作,就相当于上述排序工作的一个延续。它会在所有的文件都拷贝完毕后进行。
//使用工具类Merger归并所有的文件。经过这一个流程,一个合并了所有所需Map任务输出文件的新文件产生了。
//而那些从其他各个服务器网罗过来的 Map任务输出文件,全部删除了。 //根据hadoop是否分布式来决定调用哪种排序方式
RawKeyValueIterator rIter = isLocal
? Merger.merge(job, rfs, job.getMapOutputKeyClass(),
job.getMapOutputValueClass(), codec, getMapFiles(rfs, true),
!conf.getKeepFailedTaskFiles(), job.getInt("io.sort.factor", 100),
new Path(getTaskID().toString()), job.getOutputKeyComparator(),
reporter, spilledRecordsCounter, null)
: reduceCopier.createKVIterator(job, rfs, reporter); // free up the data structures
mapOutputFilesOnDisk.clear(); sortPhase.complete(); // sort is complete
setPhase(TaskStatus.Phase.REDUCE);
statusUpdate(umbilical);
//3.Reduce 1.Reduce任务的最后一个阶段。它会准备好Map的 keyClass("mapred.output.key.class"或"mapred.mapoutput.key.class"),
//valueClass("mapred.mapoutput.value.class"或"mapred.output.value.class")
//和 Comparator (“mapred.output.value.groupfn.class”或 “mapred.output.key.comparator.class”)
Class keyClass = job.getMapOutputKeyClass();
Class valueClass = job.getMapOutputValueClass();
RawComparator comparator = job.getOutputValueGroupingComparator();
//2.根据参数useNewAPI判断执行runNewReduce还是runOldReduce。分析润runNewReduce
if (useNewApi) {
//3.runNewReducer
//0.像报告进程书写一些信息
//1.获得一个TaskAttemptContext对象。通过这个对象创建reduce、output及用于跟踪的统计output的RecordWrit、最后创建用于收集reduce结果的Context
//2.reducer.run(reducerContext)开始执行reduce
runNewReducer(job, umbilical, reporter, rIter, comparator,
keyClass, valueClass);
} else {
runOldReducer(job, umbilical, reporter, rIter, comparator,
keyClass, valueClass);
}
done(umbilical, reporter);
}
(1)reduce分为三个阶段(copy就是远程拷贝Map的输出数据、sort就是对所有的数据做排序、reduce做聚集就是我们自己写的reducer),为这三个阶段分别设置Progress,用来和TaskTracker通信报道状态。
(2)上面代码的15-40行和MapReduce的MapTask任务的运行源码级分析 中对应部分基本相同,可参考之;
(3)codec = initCodec()这句是检查map的输出是否是压缩的,压缩的则返回压缩codec实例,否则返回null,这里讨论不压缩的;
(4)我们讨论完全分布式的hadoop,即isLocal==false,然后构造一个ReduceCopier对象reduceCopier,并调用reduceCopier.fetchOutputs()方法拷贝各个Mapper的输出,到本地;
(5)然后copy阶段完成,设置接下来的阶段是sort阶段,更新状态信息;
(6)根据isLocal来选择KV迭代器,完全分布式的会使用reduceCopier.createKVIterator(job, rfs, reporter)作为KV迭代器;
(7)sort阶段完成,设置接下来的阶段是reduce阶段,更新状态信息;
(8)然后获取一些配置信息,并根据是否使用新API选择不同的处理方式,这里是新的API,调用runNewReducer(job, umbilical, reporter, rIter, comparator, keyClass, valueClass)会执行reducer;
(9)done(umbilical, reporter)这个方法用于做结束任务的一些清理工作:更新计数器updateCounters();如果任务需要提交,设置Taks状态为COMMIT_PENDING,并利用TaskUmbilicalProtocol,汇报Task完成,等待提交,然后调用commit提交任务;设置任务结束标志位;结束Reporter通信线程;发送最后一次统计报告(通过sendLastUpdate方法);利用TaskUmbilicalProtocol报告结束状态(通过sendDone方法)。
有些人将Reduce Task分为了5个阶段:一、shuffle阶段:也称为Copy阶段,就是从各个MapTask上远程拷贝一片数据,如果大小超过一定阈值就写到磁盘,否则放入内存;二、Merge阶段:在远程拷贝数据的同时,Reduce Task启动了两个后台线程对内存和磁盘上的文件进行合并,防止内存使用过多和磁盘文件过多;三、sort阶段:用户编写的reduce方法的输入数据是按key进行聚集的,需要对copy过来的数据排序,这里用的是归并排序,因为Map Task的结果是有序的;四、Reduce阶段:将每组数据依次交给用户编写的Reduce方法处理;五、write阶段:就是将结果写入HDFS。
上面的5个阶段分的比较细了,代码里分为3个阶段copy、sort、reduce,我们在eclipse运行MR程序时,控制台看到的reduce阶段的百分比就分为3个阶段各占33.3%。
接下来重点将两个个地方:runNewReducer方法和ReduceCopier类,后者有2000多行代码,占据了ReduceTask类的绝大部分代码量。
A、我们先看runNewReducer吧,这个比ReduceCopier更容易一些,代码如下:
@SuppressWarnings("unchecked")
private <INKEY,INVALUE,OUTKEY,OUTVALUE>
void runNewReducer(JobConf job,
final TaskUmbilicalProtocol umbilical,
final TaskReporter reporter,
RawKeyValueIterator rIter,
RawComparator<INKEY> comparator,
Class<INKEY> keyClass,
Class<INVALUE> valueClass
) throws IOException,InterruptedException,
ClassNotFoundException {
// wrap value iterator to report progress.
final RawKeyValueIterator rawIter = rIter;
rIter = new RawKeyValueIterator() {
public void close() throws IOException {
rawIter.close();
}
public DataInputBuffer getKey() throws IOException {
return rawIter.getKey();
}
public Progress getProgress() {
return rawIter.getProgress();
}
public DataInputBuffer getValue() throws IOException {
return rawIter.getValue();
}
public boolean next() throws IOException {
boolean ret = rawIter.next();
reducePhase.set(rawIter.getProgress().get());
reporter.progress();
return ret;
}
};
// make a task context so we can get the classes
/*TaskAttemptContext类继承于JobContext类,相对于JobContext类增加了一些有关task的信息。通过taskContext对象可以获得很多与任务执行相
关的类,比如用户定义的Mapper类,InputFormat类等等 */
org.apache.hadoop.mapreduce.TaskAttemptContext taskContext =
new org.apache.hadoop.mapreduce.TaskAttemptContext(job, getTaskID());
// make a reducer
//创建用户定义的Reduce类的实例
org.apache.hadoop.mapreduce.Reducer<INKEY,INVALUE,OUTKEY,OUTVALUE> reducer =
(org.apache.hadoop.mapreduce.Reducer<INKEY,INVALUE,OUTKEY,OUTVALUE>)
ReflectionUtils.newInstance(taskContext.getReducerClass(), job); org.apache.hadoop.mapreduce.RecordWriter<OUTKEY,OUTVALUE> trackedRW =
new NewTrackingRecordWriter<OUTKEY, OUTVALUE>(reduceOutputCounter,
job, reporter, taskContext);
job.setBoolean("mapred.skip.on", isSkipping());
org.apache.hadoop.mapreduce.Reducer.Context
reducerContext = createReduceContext(reducer, job, getTaskID(),
rIter, reduceInputKeyCounter,
reduceInputValueCounter,
trackedRW, committer,
reporter, comparator, keyClass,
valueClass);
reducer.run(reducerContext);
trackedRW.close(reducerContext);
}
(1)参数RawKeyValueIterator rIter实际上是org.apache.hadoop.mapred.Merger.MergeQueue。这里将rIter赋值给新的RawKeyValueIterator rawIter,然后将rIter重新实现了RawKeyValueIterator,可以跟踪和汇报rawIter进度;
(2)构造任务配置类以及获取用户自己的Reducer类的实例,然后创建一个NewTrackingRecordWriter的对象trackedRW作为输出;
(3)将rIter、trackedRW等信息传递给org.apache.hadoop.mapreduce.Reducer.Context ,构造了一个管理读写的配置对象;在其父类ReduceContext中对输入就是迭代器的操作进行了实现;在ReduceContext的父类TaskInputOutputContext中实现输出的方法,其write方法会直接调用trackedRW.write(key,value)
(4)reducer.run(reducerContext)执行reducer的run方法,这个run方法和上一节中的基本相同,可参考之;
(5)关闭输出trackedRW.close(reducerContext)。
一、这里还得解释一下NewTrackingRecordWriter这个管理输出的类,是mapreduce.RecordWriter的子类,和上一节中的NewDirectOutputCollector较为类似,这里不再讲解。
二、至于输入数据rIter迭代器,在此需要解释一下,实现同一个key的不同value迭代读取的功能在ReduceContext中,讲之前,我们先看一下Reducer.run()方法的代码吧:
public void run(Context context) throws IOException, InterruptedException {
setup(context);
while (context.nextKey()) {
reduce(context.getCurrentKey(), context.getValues(), context);
}
cleanup(context);
}
我们只说while循环这一部分,其他部分前一小节有讲解,基本类似。while的循环条件是ReduceContext.nextKey()为真,这个方法就在ReduceContext中实现的,这个方法的目的就是处理下一个唯一的key(就是要保证是新的key),因为reduce方法的输入数据是分组的,所以每次都会处理一个key及这个key对应的所有value,又因为已经将所有的Map Task的输出拷贝过来而且做了排序,所以key相同的KV对都是挨着的。来看nextKey()方法代码:
/** Start processing next unique key. */
public boolean nextKey() throws IOException,InterruptedException {
while (hasMore && nextKeyIsSame) { //如果还有数据并且下一个KV中的K与当前的相同就一直循环直到key不相同,一般不会执行这个,因为value的迭代器会迭代到nextKeyIsSame==false
nextKeyValue();
}
if (hasMore) { //如果还有数据
if (inputKeyCounter != null) {
inputKeyCounter.increment(1); //统计
}
return nextKeyValue(); //推进到下一个KV
} else {
return false;
}
}
上述方法会调用另外一个方法nextKeyValue()会尝试去获取下一个key值,并且如果没数据了就会返回false,如果还有数据就返回true,具体代码如下:
public boolean nextKeyValue() throws IOException, InterruptedException {
if (!hasMore) {
key = null;
value = null;
return false;
}
firstValue = !nextKeyIsSame; //这个是否是同一个key值的不同value,第一个value的话firstValue==true并且nextKeyIsSame==false,后续的会是false,nextKeyIsSame是true
DataInputBuffer next = input.getKey();
currentRawKey.set(next.getData(), next.getPosition(),
next.getLength() - next.getPosition());
buffer.reset(currentRawKey.getBytes(), 0, currentRawKey.getLength());
key = keyDeserializer.deserialize(key); //反序列化获取key值
next = input.getValue();
buffer.reset(next.getData(), next.getPosition(), next.getLength());
value = valueDeserializer.deserialize(value); //反序列化获取value值
hasMore = input.next(); //是否还有数据
if (hasMore) {
next = input.getKey();
nextKeyIsSame = comparator.compare(currentRawKey.getBytes(), 0,
currentRawKey.getLength(),
next.getData(),
next.getPosition(),
next.getLength() - next.getPosition()
) == 0; //查看下一个KV的key是否与当前的一样
} else { //没有数据了
nextKeyIsSame = false;
}
inputValueCounter.increment(1);
return true;
}
这里面有两个比较重要的参数:firstValue表示是否是当前key值的第一个value;nextKeyIsSame表示下一个key是否和当前key值相同。这两个参数在迭代获取value的时候会有重要作用。在这个方法中会获取key和value,可以通过getCurrentKey()和getCurrentValue()方法来获取这两个值。这个方法还会读取下一个key与当前的key作比较,如果相同则nextKeyIsSame=true,否则nextKeyIsSame=false。
此时我们再返回到run()方法中,循环条件已了解,那么循环体的秘密呢?用户自己的reduce方法还记得么?一个key和一个这个key对应的value迭代器,没错在这分别对应context.getCurrentKey()和context.getValues()。下面我们重点研究一下后者context.getValues(),这个方法也在ReduceContext类中,这个方法主要是返回一个可迭代对象ValueIterable,它封装了迭代器ValueIterator,这个迭代器实现了对value的迭代读取,这个类的全部代码如下:
protected class ValueIterator implements Iterator<VALUEIN> { @Override
public boolean hasNext() {
return firstValue || nextKeyIsSame;
} @Override
public VALUEIN next() {
// if this is the first record, we don't need to advance
if (firstValue) {
firstValue = false;
return value;
}
// if this isn't the first record and the next key is different, they
// can't advance it here.
if (!nextKeyIsSame) {
throw new NoSuchElementException("iterate past last value");
}
// otherwise, go to the next key/value pair
try { //firstValue==false and nextKeyIsSame == true
nextKeyValue();
return value;
} catch (IOException ie) {
throw new RuntimeException("next value iterator failed", ie);
} catch (InterruptedException ie) {
// this is bad, but we can't modify the exception list of java.util
throw new RuntimeException("next value iterator interrupted", ie);
}
} @Override
public void remove() {
throw new UnsupportedOperationException("remove not implemented");
} }
hasNext()判断是否还有下一个value,由上面说的firstValue和nextKeyIsSame决定,只要有一个是true就说明有下一个value,为什么呢,请看上面对着两个参数的解释,自行理解吧,很明显。
next()方法就是读取value的地方,这有几种情况需要分析:1、如果firstValue==true,则直接返回当前的value,大伙这没问题吧;2、如果firstValue==false and nextKeyIsSame == false,这明显不科学,哪有下一个key不相同且又不是第一个value的情况呢?所以报错;3、如果firstValue==false and nextKeyIsSame == true 说明下一个KV的key和当前key相同且不是第一个value,可能是第N个,所以需要调用nextKeyValue()获取下一个value并返回。reduce就是通过这种机制不断去获取同一个key的所有valude的。
这个上面二中的输入数据迭代器就明了了。
B、下面就是ReduceCopier类了,这个类承载的工作量很大,也比较复杂。
重点的方法是ReduceCopier.fetchOutputs()这个方法负责拷贝各个Map函数的输出,代码也比较多接近400行,代码如下,里面有一些注释:
//通过ReduceCopier的fetchOutputs()方法取得map的结果
public boolean fetchOutputs() throws IOException {
int totalFailures = 0;
int numInFlight = 0, numCopied = 0;
DecimalFormat mbpsFormat = new DecimalFormat("0.00");
final Progress copyPhase =
reduceTask.getProgress().phase();
//(4)同时合并,还有一个内存Merger线程InMemFSMergeThread和一个文件Merger线程LocalFSMerger在同步工作,
//它们将下载过来的文件(可能在内存中,简单的统称为文件...),做着归并排序,以此,节约时间,降低输入文件的数量,
//为后续的排序工作减 负。InMemFSMergeThread的run循环调用doInMemMerge,该方法使用工具类Merger实现归并,
//如果需要combine,则combinerRunner.combine。
LocalFSMerger localFSMergerThread = null;
InMemFSMergeThread inMemFSMergeThread = null;
//(1)索取任务。使用GetMapEventsThread线程。
//该线程的run方法不停的调用getMapCompletionEvents方法,
//该方法又使用RPC调用TaskUmbilicalProtocol协议的getMapCompletionEvents,
//方法使用所属的jobID向其父TaskTracker询问此作业个Map任务 的完成状况
//(TaskTracker要向JobTracker询问后再转告给它...)。返回一个数组TaskCompletionEvent events[]。
//TaskCompletionEvent包含taskid和ip地址之类的信息。
GetMapEventsThread getMapEventsThread = null; for (int i = 0; i < numMaps; i++) {
copyPhase.addPhase(); // add sub-phase per file
} copiers = new ArrayList<MapOutputCopier>(numCopiers); // start all the copying threads
for (int i=0; i < numCopiers; i++) {
//(2)当获取到相关Map任务执行服务器的信息后,有一个线程MapOutputCopier开启,做具体的拷贝工作。
//它会在一个单独的线程内,负责某个Map任务服务器上文件的拷贝工作。MapOutputCopier的run循环调用
//copyOutput,copyOutput又调用 getMapOutput,使用HTTP远程拷贝。
MapOutputCopier copier = new MapOutputCopier(conf, reporter,
reduceTask.getJobTokenSecret());
copiers.add(copier);
copier.start();
} //start the on-disk-merge thread
localFSMergerThread = new LocalFSMerger((LocalFileSystem)localFileSys);
//start the in memory merger thread
inMemFSMergeThread = new InMemFSMergeThread();
localFSMergerThread.start();
inMemFSMergeThread.start(); // start the map events thread
getMapEventsThread = new GetMapEventsThread();
getMapEventsThread.start(); // start the clock for bandwidth measurement
long startTime = System.currentTimeMillis();
long currentTime = startTime;
long lastProgressTime = startTime;
long lastOutputTime = 0; // loop until we get all required outputs
while (copiedMapOutputs.size() < numMaps && mergeThrowable == null) { currentTime = System.currentTimeMillis();
boolean logNow = false;
if (currentTime - lastOutputTime > MIN_LOG_TIME) {
lastOutputTime = currentTime;
logNow = true;
}
if (logNow) {
LOG.info(reduceTask.getTaskID() + " Need another "
+ (numMaps - copiedMapOutputs.size()) + " map output(s) "
+ "where " + numInFlight + " is already in progress");
} // Put the hash entries for the failed fetches.
Iterator<MapOutputLocation> locItr = retryFetches.iterator(); while (locItr.hasNext()) {
MapOutputLocation loc = locItr.next();
List<MapOutputLocation> locList =
mapLocations.get(loc.getHost()); // Check if the list exists. Map output location mapping is cleared
// once the jobtracker restarts and is rebuilt from scratch.
// Note that map-output-location mapping will be recreated and hence
// we continue with the hope that we might find some locations
// from the rebuild map.
if (locList != null) {
// Add to the beginning of the list so that this map is
//tried again before the others and we can hasten the
//re-execution of this map should there be a problem
locList.add(0, loc);
}
} if (retryFetches.size() > 0) {
LOG.info(reduceTask.getTaskID() + ": " +
"Got " + retryFetches.size() +
" map-outputs from previous failures");
}
// clear the "failed" fetches hashmap
retryFetches.clear(); // now walk through the cache and schedule what we can
int numScheduled = 0;
int numDups = 0; synchronized (scheduledCopies) { // Randomize the map output locations to prevent
// all reduce-tasks swamping the same tasktracker
List<String> hostList = new ArrayList<String>();
hostList.addAll(mapLocations.keySet()); Collections.shuffle(hostList, this.random);//混洗,降低热点的出现 Iterator<String> hostsItr = hostList.iterator(); while (hostsItr.hasNext()) { String host = hostsItr.next(); List<MapOutputLocation> knownOutputsByLoc =
mapLocations.get(host); // Check if the list exists. Map output location mapping is
// cleared once the jobtracker restarts and is rebuilt from
// scratch.
// Note that map-output-location mapping will be recreated and
// hence we continue with the hope that we might find some
// locations from the rebuild map and add then for fetching.
if (knownOutputsByLoc == null || knownOutputsByLoc.size() == 0) {
continue;
} //Identify duplicate hosts here
if (uniqueHosts.contains(host)) {
numDups += knownOutputsByLoc.size();
continue;
} Long penaltyEnd = penaltyBox.get(host);
boolean penalized = false; if (penaltyEnd != null) {
if (currentTime < penaltyEnd.longValue()) {
penalized = true;
} else {
penaltyBox.remove(host);
}
} if (penalized)
continue; synchronized (knownOutputsByLoc) { locItr = knownOutputsByLoc.iterator(); while (locItr.hasNext()) { MapOutputLocation loc = locItr.next(); // Do not schedule fetches from OBSOLETE maps
if (obsoleteMapIds.contains(loc.getTaskAttemptId())) {
locItr.remove();
continue;
} uniqueHosts.add(host);
scheduledCopies.add(loc);
locItr.remove(); // remove from knownOutputs
numInFlight++; numScheduled++; break; //we have a map from this host
}
}
}
scheduledCopies.notifyAll();
} if (numScheduled > 0 || logNow) {
LOG.info(reduceTask.getTaskID() + " Scheduled " + numScheduled +
" outputs (" + penaltyBox.size() +
" slow hosts and" + numDups + " dup hosts)");
} if (penaltyBox.size() > 0 && logNow) {
LOG.info("Penalized(slow) Hosts: ");
for (String host : penaltyBox.keySet()) {
LOG.info(host + " Will be considered after: " +
((penaltyBox.get(host) - currentTime)/1000) + " seconds.");
}
} // if we have no copies in flight and we can't schedule anything
// new, just wait for a bit
try {
if (numInFlight == 0 && numScheduled == 0) {
// we should indicate progress as we don't want TT to think
// we're stuck and kill us
reporter.progress();
Thread.sleep(5000);
}
} catch (InterruptedException e) { } // IGNORE while (numInFlight > 0 && mergeThrowable == null) {
LOG.debug(reduceTask.getTaskID() + " numInFlight = " +
numInFlight);
//the call to getCopyResult will either
//1) return immediately with a null or a valid CopyResult object,
// or
//2) if the numInFlight is above maxInFlight, return with a
// CopyResult object after getting a notification from a
// fetcher thread,
//So, when getCopyResult returns null, we can be sure that
//we aren't busy enough and we should go and get more mapcompletion
//events from the tasktracker
CopyResult cr = getCopyResult(numInFlight); if (cr == null) {
break;
} if (cr.getSuccess()) { // a successful copy
numCopied++;
lastProgressTime = System.currentTimeMillis();
reduceShuffleBytes.increment(cr.getSize()); long secsSinceStart =
(System.currentTimeMillis()-startTime)/1000+1;
float mbs = ((float)reduceShuffleBytes.getCounter())/(1024*1024);
float transferRate = mbs/secsSinceStart; copyPhase.startNextPhase();
copyPhase.setStatus("copy (" + numCopied + " of " + numMaps
+ " at " +
mbpsFormat.format(transferRate) + " MB/s)"); // Note successful fetch for this mapId to invalidate
// (possibly) old fetch-failures
fetchFailedMaps.remove(cr.getLocation().getTaskId());
} else if (cr.isObsolete()) {
//ignore
LOG.info(reduceTask.getTaskID() +
" Ignoring obsolete copy result for Map Task: " +
cr.getLocation().getTaskAttemptId() + " from host: " +
cr.getHost());
} else {
retryFetches.add(cr.getLocation()); // note the failed-fetch
TaskAttemptID mapTaskId = cr.getLocation().getTaskAttemptId();
TaskID mapId = cr.getLocation().getTaskId(); totalFailures++;
Integer noFailedFetches =
mapTaskToFailedFetchesMap.get(mapTaskId);
noFailedFetches =
(noFailedFetches == null) ? 1 : (noFailedFetches + 1);
mapTaskToFailedFetchesMap.put(mapTaskId, noFailedFetches);
LOG.info("Task " + getTaskID() + ": Failed fetch #" +
noFailedFetches + " from " + mapTaskId); if (noFailedFetches >= abortFailureLimit) {
LOG.fatal(noFailedFetches + " failures downloading "
+ getTaskID() + ".");
umbilical.shuffleError(getTaskID(),
"Exceeded the abort failure limit;"
+ " bailing-out.", jvmContext);
} checkAndInformJobTracker(noFailedFetches, mapTaskId,
cr.getError().equals(CopyOutputErrorType.READ_ERROR)); // note unique failed-fetch maps
if (noFailedFetches == maxFetchFailuresBeforeReporting) {
fetchFailedMaps.add(mapId); // did we have too many unique failed-fetch maps?
// and did we fail on too many fetch attempts?
// and did we progress enough
// or did we wait for too long without any progress? // check if the reducer is healthy
boolean reducerHealthy =
(((float)totalFailures / (totalFailures + numCopied))
< MAX_ALLOWED_FAILED_FETCH_ATTEMPT_PERCENT); // check if the reducer has progressed enough
boolean reducerProgressedEnough =
(((float)numCopied / numMaps)
>= MIN_REQUIRED_PROGRESS_PERCENT); // check if the reducer is stalled for a long time
// duration for which the reducer is stalled
int stallDuration =
(int)(System.currentTimeMillis() - lastProgressTime);
// duration for which the reducer ran with progress
int shuffleProgressDuration =
(int)(lastProgressTime - startTime);
// min time the reducer should run without getting killed
int minShuffleRunDuration =
(shuffleProgressDuration > maxMapRuntime)
? shuffleProgressDuration
: maxMapRuntime;
boolean reducerStalled =
(((float)stallDuration / minShuffleRunDuration)
>= MAX_ALLOWED_STALL_TIME_PERCENT); // kill if not healthy and has insufficient progress
if ((fetchFailedMaps.size() >= maxFailedUniqueFetches ||
fetchFailedMaps.size() == (numMaps - copiedMapOutputs.size()))
&& !reducerHealthy
&& (!reducerProgressedEnough || reducerStalled)) {
LOG.fatal("Shuffle failed with too many fetch failures " +
"and insufficient progress!" +
"Killing task " + getTaskID() + ".");
umbilical.shuffleError(getTaskID(),
"Exceeded MAX_FAILED_UNIQUE_FETCHES;"
+ " bailing-out.", jvmContext);
} } currentTime = System.currentTimeMillis();
long currentBackOff = (long)(INITIAL_PENALTY *
Math.pow(PENALTY_GROWTH_RATE, noFailedFetches)); penaltyBox.put(cr.getHost(), currentTime + currentBackOff);
LOG.warn(reduceTask.getTaskID() + " adding host " +
cr.getHost() + " to penalty box, next contact in " +
(currentBackOff/1000) + " seconds");
}
uniqueHosts.remove(cr.getHost());
numInFlight--;
}
} // all done, inform the copiers to exit
exitGetMapEvents= true;
try {
getMapEventsThread.join();
LOG.info("getMapsEventsThread joined.");
} catch (InterruptedException ie) {
LOG.info("getMapsEventsThread threw an exception: " +
StringUtils.stringifyException(ie));
} synchronized (copiers) {
synchronized (scheduledCopies) {
for (MapOutputCopier copier : copiers) {
copier.interrupt();
}
copiers.clear();
}
} // copiers are done, exit and notify the waiting merge threads
synchronized (mapOutputFilesOnDisk) {
exitLocalFSMerge = true;
mapOutputFilesOnDisk.notify();
} ramManager.close(); //Do a merge of in-memory files (if there are any)
if (mergeThrowable == null) {
try {
// Wait for the on-disk merge to complete
localFSMergerThread.join();
LOG.info("Interleaved on-disk merge complete: " +
mapOutputFilesOnDisk.size() + " files left."); //wait for an ongoing merge (if it is in flight) to complete
inMemFSMergeThread.join();
LOG.info("In-memory merge complete: " +
mapOutputsFilesInMemory.size() + " files left.");
} catch (InterruptedException ie) {
LOG.warn(reduceTask.getTaskID() +
" Final merge of the inmemory files threw an exception: " +
StringUtils.stringifyException(ie));
// check if the last merge generated an error
if (mergeThrowable != null) {
mergeThrowable = ie;
}
return false;
}
}
return mergeThrowable == null && copiedMapOutputs.size() == numMaps;
}
该方法会构造多个线程对象:1个LocalFSMerger线程、1个InMemFSMergeThread线程、1个GetMapEventsThread线程、若干个(由"mapred.reduce.parallel.copies"决定,默认是5)MapOutputCopier线程。
(1)先开若干个MapOutputCopier,并启动线程,加入copiers存储列表。这个线程的run方法中有个死循环,一直监控scheduledCopies列表,这个列表表示正在拷贝的map输出的列表,当scheduledCopies一旦发现有MapOutputLocation就获取第一个MapOutputLocation,调用方法copyOutput(loc)来从远程通过HTTP拷贝Map的输出数据。copyOutput(loc)方法首先检查这个MapOutputLocation是否在copiedMapOutputs和obsoleteMapIds之中,是不能拷贝的,如果在就直接返回-2;然后通过getMapOutput(MapOutputLocation mapOutputLoc, Path filename, int reduce)方法与远程taskTracker建立连接,并获取输入流,通过一系列检查之后检查内存文件系统是否可以放得下这个map输出,如果可以放得下就通过shuffleInMemory方法将这个文件放入内存,否则通过shuffleToDisk刷新到磁盘(shuffleInMemory方法会等待内存释放足够的空间并会关闭输入流再再次建立输入流,在内存中开辟空间,将map数据拷贝到这这段空间中并封装到MapOutput中,然后返回这个MapOutput;shuffleToDisk方法首先会找一个合适的本地位置来存储map的输出,然后构造一个MapOutput对象,并从输入流持续的写到输出流指定的文件中,将这个文件封装到MapOutput中,返回MapOutput)。再返回到copyOutput方法,再对返回的MapOutput做一些检查最终如果是在内存中则mapOutputsFilesInMemory.add(mapOutput);否则是在本地磁盘对其重命名并将这个文件对应的FileStatus加入mapOutputFilesOnDisk。run方法中的finally中的finish方法将已经拷贝的MapOutputLocation放入copyResults。
(2)构造LocalFSMerger对象并启动线程,其run方法如果exitLocalFSMerge==false就会一直等待本地文件数量>=(2 * ioSortFactor - 1),会触发本地文件合并操作,ioSortFactor是参数"io.sort.factor",默认是10。然后会从 mapOutputFilesOnDisk(是SortedSet类型)中选取最小的前10个文件放入mapFiles,通过Merger.merge归并排序这10个文件,写入writer指定的文件,并将新文件放入mapOutputFilesOnDisk中。这里如果设置了combiner,也不会调用。
(3)构造InMemFSMergeThread对象并启动线程,其run方法循环检查内存中的文件是否可以合并通过exit = ramManager.waitForDataToMerge(),如果满足以下几个条件之一就会触发合并内存文件的操作:一、数据拷贝完毕后,关闭ShuffleRamManager;二、ShuffleRamManager 中已使用内存超过可用内存的“mapred.job.shuffle.merge.percent”,默认是0.66且内存文件数目超过2个;三、内存中 的文件数目超过“mapred.inmem.merge.threshold”,默认是1000;四、阻塞在ShuffleRamManager上的请求数目超过拷贝线程数"mapred.reduce.parallel.copies"的0.75。满足条件就会调用doInMemMerge()方法来执行合并操作,该方法使用工具类Merger实现归并,如果设置了combiner,则在写入本地文件之前通过combinerRunner.combine来将排序后的数据聚集后写入writer指定的本地文件中。这里有个问题要注意就是run方法中是do-while循环,循环条件是(!exit),即当exit==false时才会持续的运行,waitForDataToMerge方法中可以看出来只有ramManager关闭之后才会返回true。
(4)构造GetMapEventsThread对象并启动线程。此线程的run方法是每隔1s调用getMapCompletionEvents()方法直到exitGetMapEvents==true(会在fetchOutputs()中赋值true),这个方法会与TaskTracker通信调用TaskTracker.getMapCompletionEvents已经获取到的etionEvents方法获取已完成的Map Task列表:规则是先查找shouldReset有没有当前reduce task对应的ID,如果有说明要正在shuffle要回滚,则就返回一个要reset的MapTaskCompletionEventsUpdate;如果shouldReset没有,则从runningJobs中找到当前reduce task所属的Job的FetchStatus;获取新增的完成的map task列表FetchStatus.getMapEvents(fromEventId, maxLocs),从allMapEvents中获取需要的已完成的map,然后封装到这个列表到MapTaskCompletionEventsUpdate,再返回。那么allMapEvents中的数据是如何来的呢?TaskTracker有个MapEventsFetcherThread线程,其run方法会周期性的去获取runningJobs所有的job中第一个处于SHUFFLE阶段的reduce task对应job的FetchStatus,然后对每个FetchStatus调用其fetchMapCompletionEvents(currentTime)方法调用queryJobTracker(fromEventId, jobId, jobClient)方法与JobTracker通信通过JobTracker.getTaskCompletionEvents方法从JobInProgress中的taskCompletionEvents来获取满足条件的TaskCompletionEvent,从中找出是Map task的更新allMapEvents。
getMapCompletionEvents()方法中获取到了MapTaskCompletionEventsUpdate之后,就将已完成的map列表放入TaskCompletionEvent events[]之中;如果是reset的,则重置fromEventId、obsoleteMapIds、mapLocations;然后更新fromEventId表示已经获取到已完成map的最新编号,以后再获取新增将会是这个编号之后的。然后遍历events中的所有TaskCompletionEvent,根据每个的状态:如果是SUCCEEDED,则放入mapLocations(保存了TaskTracker Host与已完成任务列表的映射关系)可以去取map的输出数据;如果是OBSOLETE/FAILED/KILLED,就放入obsoleteOutputs,表示停止从这些map取数据;如果是TIPFAILED,则放入copiedMapOutputs表示不需要从这些map去取数据。然后返回mapLocations新增的的个数。
在fetchOutputs()方法中这些线程启动之后,还不能工作,还需要将mapLocations中合适的MapOutputLocation放入scheduledCopies唤醒MapOutputCopier线程去拷贝,如果A、所有的拷贝结果中会将拷贝成功的从fetchFailedMaps中删除;B、是Obsolete的会忽略;C、其他失败的加入retryFetches,并且对应mapTaskId的失败次数会加1,并放入mapTaskToFailedFetchesMap之中,这个结构是用来存放mapTaskId和对应的失败次数的,容错机制一:拷贝失败次数超过上限(Math.max(30, numMaps / 10))就会杀死该Reduce Task(等待调度器重新调度执行);容错机制二:一旦拷贝失败次数>=maxFetchFailuresBeforeReporting(由参数"mapreduce.reduce.shuffle.maxfetchfailures"指定,默认是10),就加入fetchFailedMaps,同时满足以下条件就会杀死这个reduce task:一、reducer所在节点不健康;二、fetchFailedMaps的大小超过上限(默认是5)或者等于所有的reducer需要的所有的map的个数减去copiedMapOutputs的大小;三、reducer没有足够的Progress或者reducer超时停滞了,容错三、如果前两个条件均不满足,则采用对数回归模型推迟一段时间后重新拷贝对应的map的输出数据,延迟时间是10000*Math.pow(1.3, noFailedFetches)),并放入penaltyBox中进行惩罚。最后待copy操作完成会做一些清理工作:会关闭ramManager,触发InMemFSMergeThread线程结束退出;exitGetMapEvents=true会使得GetMapEventsThread结束退出;exitLocalFSMerge=true会使得LocalFSMerger线程结束退出;挨个中断copiers中所有拷贝线程MapOutputCopier,清理copiers.clear()。
至此reduce task算是讲解完毕,mapreduce的整个过程已经讲解了很多内容,大体的过程已知。还有许多东西没有涉及,比如恢复机制、容错机制、任务的推测、快排和归并、文件流的过程包括文件名和位置等等。后续还会继续研究。
参考:1、董西成,《hadoop技术内幕---深入理解MapReduce架构设计与实现原理》