In this cluster, data nodes runs as user 'mapred'. Actually, all hadoop daemons runs as user 'mapred'.
2013/12/22 Ted Yu <[email protected]> > Are your data nodes running as user 'hdfs', or 'mapred' ? > > If the former, you need to increase file limit for 'hdfs' user. > > Cheers > > > On Sat, Dec 21, 2013 at 8:30 AM, sam liu <[email protected]> wrote: > >> Hi Experts, >> >> We failed to run an MR job which accesses hive, as hdfs is unable to >> create new block during reduce phase. The exceptions: >> 1) In tasklog: >> hdfs.DFSClient: DataStreamer Exception: java.io.IOException: Unable to >> create new block >> 2) In HDFS data node log: >> DataXceiveServer: IOException due to:java.io.IOException: Too many open >> fiiles >> ... ... >> at sun.nio.ch.ServerSocketAdapter.accept(ServerSocketAdaptor.java:96) >> at >> org.apache.hadoop.hdfs.server.datanode.DataXceiverServer.run(DataXceiverServer.java:131) >> >> In hdfs-site.xml, we set 'dfs.datanode.max.xcievers' to 8196. At the same >> time, we modified /etc/security/limits.conf to increase nofile of mapred >> user to 1048576. But this issue still happen. >> >> Any suggestions? >> >> Thanks a lot! >> >> >
