[HDFS-7539] Namenode can't leave safemode because of Datanodes' IPC socket timeout

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.5.1
• Component/s: datanode, namenode

Environment

1 master, 1 secondary and 128 slaves, each node has x24 cores, 48GB memory. fsimage is 4GB.

• Assignee: Unassigned
• Reporter: hoelog
• Created: 17/Dec/14 06:48
• Updated: 22/Dec/14 22:12

Description

During the starting of namenode, data nodes seem waiting namenode's response through IPC to register block pools.

here is DN's log -

 
2014-12-16 20:28:09,064 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Acknowledging ACTIVE Namenode Block pool BP-877672386-10.114.130.143-1412666752827 (Datanode Uuid 2117395f-e034-4b4a-adec-8a28464f4796) service to NN.x.com/10.x.x143:9000 

But namenode is too busy to responde it, and datanodes occur socket timeout - default is 1 minute.

2014-12-16 20:29:09,857 WARN org.apache.hadoop.hdfs.server.datanode.DataNode: IOException in offerService
java.net.SocketTimeoutException: Call From DN1.x.com/10.x.x.84 to NN.x.com:9000 failed on socket timeout exception: java.net.SocketTimeoutException: 60000 millis timeout while waiting for channel to be ready for read. ch : java.nio.channels.SocketChannel[connected local=/10.x.x.84:57924 remote=NN.x.com/10.x.x.143:9000]; For more details see:  http://wiki.apache.org/hadoop/SocketTimeout 

same events repeat and eventually NN drops most connecting trials from DN. So NN can't leave safemode.

DN's log -

2014-12-16 20:32:25,895 WARN org.apache.hadoop.hdfs.server.datanode.DataNode: IOException in offerService
java.io.IOException: failed on local exception java.io.ioexception connection reset by peer

There is no troubles in the network, configuration or servers. I think NN is too busy to respond to DN in a minute.

I configured "ipc.ping.interval" to 15 mins In the core-site.xml, and that was helpful for my cluster.

<property>
  <name>ipc.ping.interval</name>
  <value>900000</value>
</property>

In my cluster, namenode responded 1 min ~ 5 mins for the DNs' request.
It will be helpful if there is more elegant solution.

2014-12-16 23:28:16,598 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Acknowledging ACTIVE Namenode Block pool BP-877672386-10.x.x.143-1412666752827 (Datanode Uuid c4f7beea-b8e9-404f-bc81-6e87e37263d2) service to NN/10.x.x.143:9000
2014-12-16 23:31:32,026 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Sent 1 blockreports 2090961 blocks total. Took 1690 msec to generate and 193738 msecs for RPC and NN processing.  Got back commands org.apache.hadoop.hdfs.server.protocol.FinalizeCommand@20e68e11
2014-12-16 23:31:32,026 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Got finalize command for block pool BP-877672386-10.x.x.143-1412666752827
2014-12-16 23:31:32,032 INFO org.apache.hadoop.util.GSet: Computing capacity for map BlockMap
2014-12-16 23:31:32,032 INFO org.apache.hadoop.util.GSet: VM type       = 64-bit
2014-12-16 23:31:32,044 INFO org.apache.hadoop.util.GSet: 0.5% max memory 3.6 GB = 18.2 MB
2014-12-16 23:31:32,045 INFO org.apache.hadoop.util.GSet: capacity      = 2^21 = 2097152 entries
2014-12-16 23:31:32,046 INFO org.apache.hadoop.hdfs.server.datanode.BlockPoolSliceScanner: Periodic Block Verification Scanner initialized with interval 504 hours for block pool BP-877672386-10.114.130.143-1412666752827