I think heartbeat failure cause is hang of nodes. I found a bug report associated with this problem. https://issues.apache.org/jira/browse/HDFS-7489 https://issues.apache.org/jira/browse/HDFS-7496 https://issues.apache.org/jira/browse/HDFS-7531 https://issues.apache.org/jira/browse/HDFS-8051 It has been fixed in 2.7. I do not have experience patch. And Because of this stability has not been confirmed, I can not upgrade to 2.7. What do you recommend for that? How can I do the patch, if I will do patch? Can I patch without service dowtime. -----Original Message----- From: "Drake민영근"<[email protected]> To: "user"<[email protected]>; "조주일"<[email protected]>; Cc: Sent: 2015-04-24 (금) 17:41:59 Subject: Re: rolling upgrade(2.4.1 to 2.6.0) problem Hi, I think limited by "max user processes". see this: https://plumbr.eu/outofmemoryerror/unable-to-create-new-native-thread In your case, user cannot create more than 10240 processes. In our env, the limit is more like "65000". I think it's worth a try. And, if hdfs datanode daemon's user is not root, set the limit file into /etc/security/limits.d Thanks.Drake 민영근 Ph.Dkt NexR
On Fri, Apr 24, 2015 at 5:15 PM, 조주일 <[email protected]> wrote: ulimit -a core file size (blocks, -c) 0 data seg size (kbytes, -d) unlimited scheduling priority (-e) 0 file size (blocks, -f) unlimited pending signals (-i) 62580 max locked memory (kbytes, -l) 64 max memory size (kbytes, -m) unlimited open files (-n) 102400 pipe size (512 bytes, -p) 8 POSIX message queues (bytes, -q) 819200 real-time priority (-r) 0 stack size (kbytes, -s) 10240 cpu time (seconds, -t) unlimited max user processes (-u) 10240 virtual memory (kbytes, -v) unlimited file locks (-x) unlimited ------------------------------------------------------ Hadoop cluster was operating normally in the 2.4.1 version. Hadoop cluster is a problem in version 2.6. E.g Slow BlockReceiver logs are often seen "org.apache.hadoop.hdfs.server.datanode.DataNode: Slow BlockReceiver write data to disk cost" If the data node failure and under-block occurs, another many nodes heartbeat check is fails. So, I stop all nodes and I start all nodes. The cluster is then normalized. In this regard, Hadoop Is there a difference between version 2.4 and 2.6? ulimit -a core file size (blocks, -c) 0 data seg size (kbytes, -d) unlimited scheduling priority (-e) 0 file size (blocks, -f) unlimited pending signals (-i) 62580 max locked memory (kbytes, -l) 64 max memory size (kbytes, -m) unlimited open files (-n) 102400 pipe size (512 bytes, -p) 8 POSIX message queues (bytes, -q) 819200 real-time priority (-r) 0 stack size (kbytes, -s) 10240 cpu time (seconds, -t) unlimited max user processes (-u) 10240 virtual memory (kbytes, -v) unlimited file locks (-x) unlimited -----Original Message----- From: "Drake민영근"<[email protected]> To: "user"<[email protected]>; "조주일"<[email protected]>; Cc: Sent: 2015-04-24 (금) 16:58:46 Subject: Re: rolling upgrade(2.4.1 to 2.6.0) problem HI, How about the ulimit setting of the user for hdfs datanode ?Drake 민영근 Ph.Dkt NexR On Wed, Apr 22, 2015 at 6:25 PM, 조주일 <[email protected]> wrote: I allocated 5G. I think OOM is not the cause of essentially -----Original Message----- From: "Han-Cheol Cho"<[email protected]> To: <[email protected]>; Cc: Sent: 2015-04-22 (수) 15:32:35 Subject: RE: rolling upgrade(2.4.1 to 2.6.0) problem Hi, The first warning shows out-of-memory error of JVM. Did you give enough max heap memory for DataNode daemons? DN daemons, by default, uses max heap size 1GB. So if your DN requires more than that, it will be in a trouble. You can check the memory consumption of you DN dameons (e.g., top command) and the memory allocated to them by -Xmx option (e.g., jps -lmv). If the max heap size is too small, you can use HADOOP_DATANODE_OPTS variable (e.g., HADOOP_DATANODE_OPTS="-Xmx4g") to override it. Best wishes, Han-Cheol -----Original Message----- From: "조주일"<[email protected]> To: <[email protected]>; Cc: Sent: 2015-04-22 (수) 14:54:16 Subject: rolling upgrade(2.4.1 to 2.6.0) problem My Cluster is.. hadoop 2.4.1 Capacity : 1.24PB Used 1.1PB 16 Datanodes Each node is a capacity of 65TB, 96TB, 80TB, Etc.. I had to proceed with the rolling upgrade 2.4.1 to 2.6.0. A data node upgraded takes about 40 minutes. Occurs during the upgrade is in progress under-block. 10 nodes completed upgrade 2.6.0. Had a problem at some point during a rolling upgrade of the remaining nodes. Heartbeat of the many nodes(2.6.0 only) has failed. I did changes the following attributes but I did not fix the problem, dfs.datanode.handler.count = 100 ---> 300, 400, 500 dfs.datanode.max.transfer.threads = 4096 ---> 8000, 10000 I think, 1. Something that causes a delay in processing threads. I think it may be because the block replication between different versions. 2. Whereby the many handlers and xceiver became necessary. 3. Whereby the out of memory, an error occurs. Or the problem arises on a datanode. 4. Heartbeat fails, and datanode dies. I found a datanode error log for the following: However, it is impossible to determine the cause. I think, therefore I am. Called because it blocks the replication between different versions Give me someone help me !! DATANODE LOG -------------------------------------------------------------------------- ### I had to check a few thousand close_wait connection from the datanode. org.apache.hadoop.hdfs.server.datanode.DataNode: Slow BlockReceiver write packet to mirror took 1207ms (threshold=300ms) 2015-04-21 22:46:01,772 WARN org.apache.hadoop.hdfs.server.datanode.DataNode: DataNode is out of memory. Will retry in 30 seconds. java.lang.OutOfMemoryError: unable to create new native thread at java.lang.Thread.start0(Native Method) at java.lang.Thread.start(Thread.java:640) at org.apache.hadoop.hdfs.server.datanode.DataXceiverServer.run(DataXceiverServer.java:145) at java.lang.Thread.run(Thread.java:662) 2015-04-21 22:49:45,378 WARN org.apache.hadoop.hdfs.server.datanode.DataNode: datanode-192.168.1.207:40010:DataXceiverServer:java.io.IOException: Xceiver count 8193 exceeds the limit of concurrent xcievers: 8192 at org.apache.hadoop.hdfs.server.datanode.DataXceiverServer.run(DataXceiverServer.java:140) at java.lang.Thread.run(Thread.java:662) 2015-04-22 01:01:25,632 WARN org.apache.hadoop.hdfs.server.datanode.DataNode: datanode-192.168.1.207:40010:DataXceiverServer:java.io.IOException: Xceiver count 8193 exceeds the limit of concurrent xcievers: 8192 at org.apache.hadoop.hdfs.server.datanode.DataXceiverServer.run(DataXceiverServer.java:140) at java.lang.Thread.run(Thread.java:662) 2015-04-22 03:49:44,125 ERROR org.apache.hadoop.hdfs.server.datanode.DataNode: datanode-192.168.1.204:40010:DataXceiver error processing READ_BLOCK operation src: /192.168.2.174:45606 dst: /192.168.1.204:40010 java.io.IOException: cannot find BPOfferService for bpid=BP-1770955034-0.0.0.0-1401163460236 at org.apache.hadoop.hdfs.server.datanode.DataNode.getDNRegistrationForBP(DataNode.java:1387) at org.apache.hadoop.hdfs.server.datanode.DataXceiver.readBlock(DataXceiver.java:470) at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.opReadBlock(Receiver.java:116) at org.apache.hadoop.hdfs.protocol.datatransfer.Receiver.processOp(Receiver.java:71) at org.apache.hadoop.hdfs.server.datanode.DataXceiver.run(DataXceiver.java:235) at java.lang.Thread.run(Thread.java:662) 2015-04-22 05:30:28,947 WARN org.apache.hadoop.hdfs.server.datanode.DataNode: DatanodeRegistration(192.168.1.203, datanodeUuid=654f22ef-84b3-4ecb-a959-2ea46d817c19, infoPort=40075, ipcPort=40020, storageInfo=lv=-56;cid=CID-CLUSTER;nsid=239138164;c=1404883838982):Failed to transfer BP-1770955034-0.0.0.0-1401163460236:blk_1075354042_1613403 to 192.168.2.156:40010 got java.net.SocketException: Original Exception : java.io.IOException: Connection reset by peer at sun.nio.ch.FileChannelImpl.transferTo0(Native Method) at sun.nio.ch.FileChannelImpl.transferToDirectly(FileChannelImpl.java:405) at sun.nio.ch.FileChannelImpl.transferTo(FileChannelImpl.java:506) at org.apache.hadoop.net.SocketOutputStream.transferToFully(SocketOutputStream.java:223) at org.apache.hadoop.hdfs.server.datanode.BlockSender.sendPacket(BlockSender.java:559) at org.apache.hadoop.hdfs.server.datanode.BlockSender.sendBlock(BlockSender.java:728) at org.apache.hadoop.hdfs.server.datanode.DataNode$DataTransfer.run(DataNode.java:2017) at java.lang.Thread.run(Thread.java:662) Caused by: java.io.IOException: Connection reset by peer ... 8 more
