HDFS does not work very well on multi-homed machines. A few open Jiras refer to this:
There are multiple issues involved here and some of them can be worked around by using alternate DNS names and configuring slave.host.name on Datanodes and task trackers.
However namenode issues cannot be worked around because it does not respect the fs.default.name configuration. e.g. Namenode#initialize performs a gratuitous reverse DNS lookup to regenerate the hostname. Similar issues exist elsewhere.
This Jira is being filed to fix some of the more serious problems. To avoid affecting existing users a new config setting may be introduced.