[Hadoop-common-user] hadoop on unstable nodes

Condor has a hadoop subproject in UW-Madison, and there are also some
scientists from VT. They worked on security Hadoop MapReduce on Internet.

In my opinion, Alex is correct, Hadoop MR is communication intensive
especially in the map and shuffle stage. In the map stage, every mapper
needs input data from File System. If your data distributed among Internet,
you may encounter heavy delay. Also in the shuffle stage, reducer collect
mapper's intermediate results through Internet. This is another bottleneck
we can not overlook.

Hope this will help.

Chen