From http://hadoop.apache.org/common/docs/current/mapred-default.html about mapred.child.ulimit:
"The maximum virtual memory, in KB, of a process launched by the Map-Reduce framework. This can be used to control both the Mapper/Reducer tasks and applications using Hadoop Pipes, Hadoop Streaming etc. By default it is left unspecified to let cluster admins control it via limits.conf and other such relevant mechanisms. Note: mapred.child.ulimit must be greater than or equal to the -Xmx passed to JavaVM, else the VM might not start."
Tom it looks like "unlimited" is not a valid value for mapred.child.ulimit. Any suggestions?