Because bin/hadoop and bin/hdfs put the entire CLASSPATH in the command line arguments, it exceeds 4096 bytes, which is the maximum size that ps (or /proc) can work with. This makes looking for the processes difficult, since the output gets truncated for all components at the same point (e.g. NameNode, SecondaryNameNode, DataNode).
The mapred sub-project does not have this problem, because it calls "export CLASSPATH" before the final exec. bin/hadoop and bin/hdfs should do the same thing