DFSClient频繁发生GC，导致写入hdfs超时的问题分析
搜索推荐有一个job，1000多个map，200个reduce，运行到最后只剩一个reduce（10.39.6.130上）的时候，出现以下异常，导致job失败：
2014-12-04 15:49:04,297 INFO [main] org.apache.hadoop.mapred.Merger: Down to the last merge-pass, with 12 segments left of total size: 11503294914 bytes
2014-12-04 15:49:04,314 INFO [main] org.apache.hadoop.conf.Configuration.deprecation: mapred.task.id is deprecated. Instead, use mapreduce.task.attempt.id
2014-12-04 15:49:04,394 INFO [main] org.apache.hadoop.io.compress.CodecPool: Got brand-new compressor [.lzo_deflate]
2014-12-04 16:02:26,889 WARN [ResponseProcessor for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086] org.apache.hadoop.hdfs.DFSClient: DFSOutputStream ResponseProcessor exception for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086
java.io.IOException: Bad response ERROR for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086 from datanode 10.39.5.193:50010
at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer$ResponseProcessor.run(DFSOutputStream.java:819)
2014-12-04 16:02:26,889 WARN [ResponseProcessor for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223] org.apache.hadoop.hdfs.DFSClient: DFSOutputStream ResponseProcessor exception for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223
java.io.IOException: Bad response ERROR for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223 from datanode 10.39.1.90:50010
at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer$ResponseProcessor.run(DFSOutputStream.java:819)
2014-12-04 16:02:26,891 WARN [DataStreamer for file /dw_ext/recmd/mds6/mds_filter_relation_10/20141203/_temporary/1/_temporary/attempt_1415948652989_195149_r_000158_3/user-r-00158 block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086] org.apache.hadoop.hdfs.DFSClient: Error Recovery for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086 in pipeline 10.39.6.130:50010, 10.39.5.185:50010, 10.39.5.193:50010: bad datanode 10.39.5.193:50010
2014-12-04 16:02:26,891 WARN [DataStreamer for file /dw_ext/recmd/mds6/mds_filter_relation_10/20141203/_temporary/1/_temporary/attempt_1415948652989_195149_r_000158_3/exposure-r-00158 block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223] org.apache.hadoop.hdfs.DFSClient: Error Recovery for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223 in pipeline 10.39.6.130:50010, 10.39.1.89:50010, 10.39.1.90:50010: bad datanode 10.39.1.90:50010
java.io.EOFException: Premature EOF: no length prefix available
at org.apache.hadoop.hdfs.protocolPB.PBHelper.vintPrefixed(PBHelper.java:1987)
at org.apache.hadoop.hdfs.protocol.datatransfer.PipelineAck.readFields(PipelineAck.java:176)
at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer$ResponseProcessor.run(DFSOutputStream.java:796)
2014-12-04 16:05:23,743 FATAL [main] org.apache.hadoop.mapred.YarnChild: Error running child : java.lang.OutOfMemoryError: Java heap space
at java.util.Arrays.copyOf(Arrays.java:2734)
at java.util.Vector.ensureCapacityHelper(Vector.java:226)
at java.util.Vector.add(Vector.java:728)
at rec.CommonUtil.pack_Treeset(CommonUtil.java:395)
at rec.ConvertExposure10$MyReducer.collect_exposure(ConvertExposure10.java:259)
at rec.ConvertExposure10$MyReducer.reduce(ConvertExposure10.java:329)
at rec.ConvertExposure10$MyReducer.reduce(ConvertExposure10.java:234)
at org.apache.hadoop.mapred.ReduceTask.runOldReducer(ReduceTask.java:444)
at org.apache.hadoop.mapred.ReduceTask.run(ReduceTask.java:392)
at org.apache.hadoop.mapred.YarnChild$2.run(YarnChild.java:167)
at java.security.AccessController.doPrivileged(Native Method)
at javax.security.auth.Subject.doAs(Subject.java:396)
at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1550)
at org.apache.hadoop.mapred.YarnChild.main(YarnChild.java:162)
从异常上看，首先是reduce在往hdfs写数据时，发现建pipeline时，没有收到pipeline上最后一个节点的回应：
2014-12-04 16:02:26,889 WARN [ResponseProcessor for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086] org.apache.hadoop.hdfs.DFSClient: DFSOutputStream ResponseProcessor exception for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086
java.io.IOException: Bad response ERROR for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086 from datanode 10.39.5.193:50010
at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer$ResponseProcessor.run(DFSOutputStream.java:819)
2014-12-04 16:02:26,889 WARN [ResponseProcessor for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223] org.apache.hadoop.hdfs.DFSClient: DFSOutputStream ResponseProcessor exception for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223
java.io.IOException: Bad response ERROR for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223 from datanode 10.39.1.90:50010
at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer$ResponseProcessor.run(DFSOutputStream.java:819)
2014-12-04 16:02:26,891 WARN [DataStreamer for file /dw_ext/recmd/mds6/mds_filter_relation_10/20141203/_temporary/1/_temporary/attempt_1415948652989_195149_r_000158_3/user-r-00158 block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086] org.apache.hadoop.hdfs.DFSClient: Error Recovery for block BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086 in pipeline 10.39.6.130:50010, 10.39.5.185:50010, 10.39.5.193:50010: bad datanode 10.39.5.193:50010
2014-12-04 16:02:26,891 WARN [DataStreamer for file /dw_ext/recmd/mds6/mds_filter_relation_10/20141203/_temporary/1/_temporary/attempt_1415948652989_195149_r_000158_3/exposure-r-00158 block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223] org.apache.hadoop.hdfs.DFSClient: Error Recovery for block BP-1386326728-10.39.2.131-1382089338395:blk_1394153869_320473223 in pipeline 10.39.6.130:50010, 10.39.1.89:50010, 10.39.1.90:50010: bad datanode 10.39.1.90:50010
java.io.EOFException: Premature EOF: no length prefix available
at org.apache.hadoop.hdfs.protocolPB.PBHelper.vintPrefixed(PBHelper.java:1987)
at org.apache.hadoop.hdfs.protocol.datatransfer.PipelineAck.readFields(PipelineAck.java:176)
at org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer$ResponseProcessor.run(DFSOutputStream.java:796)
这里以写block：blk_1394149732_320469086为例，pipeline[10.39.6.130:50010, 10.39.5.185:50010, 10.39.5.193:50010]上面的最后一个DN是10.39.5.193，到10.39.5.193查看该block的日志信息：
2014-12-04 16:00:57,424 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Exception for BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086
java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/10.39.5.193:50010 remote=/10.39.5.185:58225]
at org.apache.hadoop.net.SocketIOWithTimeout.doIO(SocketIOWithTimeout.java:164)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:161)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:131)
at java.io.BufferedInputStream.fill(BufferedInputStream.java:218)
at java.io.BufferedInputStream.read1(BufferedInputStream.java:258)
at java.io.BufferedInputStream.read(BufferedInputStream.java:317)
at java.io.DataInputStream.read(DataInputStream.java:132)
at org.apache.hadoop.io.IOUtils.readFully(IOUtils.java:192)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doReadFully(PacketReceiver.java:213)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doRead(PacketReceiver.java:134)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.receiveNextPacket(PacketReceiver.java:109)
at org.apache.hadoop.hdfs.server.datanode.BlockReceiver.receivePacket(BlockReceiver.java:446)
at org.apache.hadoop.hdfs.server.datanode.BlockReceiver.receiveBlock(BlockReceiver.java:702)
at org.apache.hadoop.hdfs.server.datanode.DataXceiver.writeBlock(DataXceiver.java:739)
at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.opWriteBlock(Receiver.java:124)
at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.processOp(Receiver.java:71)
at org.apache.hadoop.hdfs.server.datanode.DataXceiver.run(DataXceiver.java:232)
at java.lang.Thread.run(Thread.java:662)
2014-12-04 16:00:57,424 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: PacketResponder: BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086, type=LAST_IN_PIPELINE, downstreams=0:[]: Thread is interrupted.
10.39.5.193上面日志显示，在读取pipeline上一个节点10.39.5.185的Packet时，一直读取不到，直到10分钟超时：
java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/10.39.5.193:50010 remote=/10.39.5.185:58225]
那我们来看以下pipeline上第二个节点10.39.5.185，dn日志如下：
2014-12-04 16:00:57,988 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: PacketResponder: BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086, type=HAS_DOWNSTREAM_IN_PIPELINE
java.io.EOFException: Premature EOF: no length prefix available
at org.apache.hadoop.hdfs.protocolPB.PBHelper.vintPrefixed(PBHelper.java:1987)
at org.apache.hadoop.hdfs.protocol.datatransfer.PipelineAck.readFields(PipelineAck.java:176)
at org.apache.hadoop.hdfs.server.datanode.BlockReceiver$PacketResponder.run(BlockReceiver.java:1083)
at java.lang.Thread.run(Thread.java:662)
2014-12-04 16:00:58,008 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Exception for BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086
java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/10.39.5.185:50010 remote=/10.39.6.130:59083]
at org.apache.hadoop.net.SocketIOWithTimeout.doIO(SocketIOWithTimeout.java:164)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:161)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:131)
at java.io.BufferedInputStream.fill(BufferedInputStream.java:218)
at java.io.BufferedInputStream.read1(BufferedInputStream.java:258)
at java.io.BufferedInputStream.read(BufferedInputStream.java:317)
at java.io.DataInputStream.read(DataInputStream.java:132)
at org.apache.hadoop.io.IOUtils.readFully(IOUtils.java:192)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doReadFully(PacketReceiver.java:213)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doRead(PacketReceiver.java:134)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.receiveNextPacket(PacketReceiver.java:109)
at org.apache.hadoop.hdfs.server.datanode.BlockReceiver.receivePacket(BlockReceiver.java:446)
at org.apache.hadoop.hdfs.server.datanode.BlockReceiver.receiveBlock(BlockReceiver.java:702)
at org.apache.hadoop.hdfs.server.datanode.DataXceiver.writeBlock(DataXceiver.java:739)
at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.opWriteBlock(Receiver.java:124)
at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.processOp(Receiver.java:71)
at org.apache.hadoop.hdfs.server.datanode.DataXceiver.run(DataXceiver.java:232)
at java.lang.Thread.run(Thread.java:662)
2014-12-04 16:00:58,009 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: PacketResponder: BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086, type=HAS_DOWNSTREAM_IN_PIPELINE: Thread is interrupted.
和10.39.5.193日志类似，也是在等待读取pipeline的第一个节点10.39.6.130的Packet时超时：
java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/10.39.5.185:50010 remote=/10.39.6.130:59083]
这样说来，问题出在10.39.6.130上，也即reduce任务运行的节点上，该节点DN日志如下：
2014-12-04 16:00:59,987 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Exception for BP-1386326728-10.39.2.131-1382089338395:blk_1394149732_320469086
java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/10.39.6.130:50010 remote=/10.39.6.130:45259]
at org.apache.hadoop.net.SocketIOWithTimeout.doIO(SocketIOWithTimeout.java:164)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:161)
at org.apache.hadoop.net.SocketInputStream.read(SocketInputStream.java:131)
at java.io.BufferedInputStream.fill(BufferedInputStream.java:218)
at java.io.BufferedInputStream.read1(BufferedInputStream.java:258)
at java.io.BufferedInputStream.read(BufferedInputStream.java:317)
at java.io.DataInputStream.read(DataInputStream.java:132)
at org.apache.hadoop.io.IOUtils.readFully(IOUtils.java:192)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doReadFully(PacketReceiver.java:213)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.doRead(PacketReceiver.java:134)
at org.apache.hadoop.hdfs.protocol.datatransfer.PacketReceiver.receiveNextPacket(PacketReceiver.java:109)
at org.apache.hadoop.hdfs.server.datanode.BlockReceiver.receivePacket(BlockReceiver.java:446)
at org.apache.hadoop.hdfs.server.datanode.BlockReceiver.receiveBlock(BlockReceiver.java:702)
at org.apache.hadoop.hdfs.server.datanode.DataXceiver.writeBlock(DataXceiver.java:739)
at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.opWriteBlock(Receiver.java:124)
at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.processOp(Receiver.java:71)
at org.apache.hadoop.hdfs.server.datanode.DataXceiver.run(DataXceiver.java:232)
at java.lang.Thread.run(Thread.java:662)
但是根据日志信息，10.39.6.130的DN也是在等待Packet，但是一直等到超时也没等到：
java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/10.39.6.130:50010 remote=/10.39.6.130:45259]
看来不是10.39.6.130节点DN的问题，如果pipeline上面三个DN都没有问题，那问题肯定就出在dfsclient端了，也就是reduce任务在往hdfs写数据的时候根本就没有写得出去，在dfsclient上面就给堵住了，接下来查看dfsclient，也就是reduce任务进程的执行情况：
在10.39.6.130上，根据任务id：attempt_1415948652989_195149_r_000158_3 找到进程id 31050，查看内存使用情况：
jstat -gcutil 31050 1000：
S0 S1 E O P YGC YGCT FGC FGCT GCT
68.95 0.00 0.00 92.98 66.32 111 16.825 10 25.419 42.244
68.95 0.00 0.00 92.98 66.32 111 16.825 10 25.419 42.244
68.95 0.00 0.00 92.98 66.32 111 16.825 10 25.419 42.244
68.95 0.00 0.00 92.98 66.32 111 16.825 10 25.419 42.244
68.95 0.00 0.00 92.98 66.32 111 16.825 10 25.419 42.244
68.95 0.00 0.00 92.98 66.32 111 16.825 10 25.419 42.244
68.95 0.00 26.75 100.00 66.32 111 16.825 10 25.419 42.244
0.00 0.00 31.85 100.00 68.16 111 16.825 10 44.767 61.591
0.00 0.00 35.37 100.00 68.16 111 16.825 10 44.767 61.591
0.00 0.00 40.64 100.00 68.16 111 16.825 10 44.767 61.591
0.00 0.00 45.35 100.00 68.16 111 16.825 10 44.767 61.591
0.00 0.00 48.87 100.00 68.16 111 16.825 10 44.767 61.591
0.00 0.00 54.14 100.00 68.16 111 16.825 10 44.767 61.591
0.00 0.00 58.85 100.00 68.16 111 16.825 10 44.767 61.591
果然，JVM老年代被占满，不断执行FGC，直接stop the world，导致JVM没法对外提供服务，导致dfsclient挂起，一直不能往pipeline上面的节点写Packet，直到socket超时。
根据reduce任务最后的日志，也从侧面验证了这种观点：
Error running child : java.lang.OutOfMemoryError: Java heap space
既然是OOM导致的job失败，那是什么对象导致的内存泄露呢：
执行：
jmap -histo:live 31050 > jmap.log
cat jmap.log :
num #instances #bytes class name
----------------------------------------------
1: 71824177 2872967080 java.util.TreeMap$Entry
2: 71822939 1723750536 java.lang.Long
3: 10684 24777776 [B
4: 47174 6425152
5: 47174 6408120
6: 3712 4429776
7: 66100 3979224
8: 3712 2938192
9: 3125 2562728
10: 3477 1267752 [I
11: 12923 1180224 [C
12: 1794 772488
13: 13379 428128 java.lang.String
14: 4034 419536 java.lang.Class
15: 6234 410312 [S
16: 6409 352576 [[I
17: 7567 242144 java.util.HashMap$Entry
18: 293 171112
19: 4655 148960 java.util.Hashtable$Entry
20: 1535 135080 java.lang.reflect.Method
21: 842 121696 [Ljava.util.HashMap$Entry;
果然啊，reduce代码中使用了TreeMap，往里面放置了大量对象，导致出现OOM，TreeMap的Entry就站用了2.8G内存，而我们reduce设置的内存也就1.5G。
总结：对该job出现的异常，一般在以下几种情况下发生：
1、写数据块的DN出现问题，不能写入，就像之前出现的DN由于本地读问题导致xceivers(每个DN用于并发数据传输处理最大线程数)达到4096，耗尽了所有的线程，没法对新发起的输出写入做出相应。
2、网络出现异常，DN节点进或出的带宽被耗尽，导致数据写不出去或者写不进来，这种情况可以观察ganglia看节点带宽使用情况，这种情况一般比较少。当该job出现问题的时候，也怀疑过是带宽被耗尽的问题，查看了一下相关节点ganglia带宽使用情况，最大in/out 85M/s左右，最后排除是带宽问题。
3、dfsclient出现问题，长时间没反应，导致已经发起的socket超时。由于dfsclient情况比较复杂，出现问题的情况比较多，比如本问就是因为reduce出现内存溢出，jvm不断进行FGC，导致dfsclient挂起，最终socket出现超时。