Dan Templeton, who I used to work with at the Grid Forum sessions, has written up on Sun Grid Engine's Hadoop integration

A key feature is they've added location awareness to the scheduler; something runs on each datanode to log and report the blocks there, so when work is scheduled against that data, it can be run on or near that node.

There are some devious things going on there, I suspect. What's interesting here is you can do other kinds of work than just MapReduce, and when you do want to work against data -be it for MapReduce or other tasks- then the scheduler can do work near it. This makes it an alternative to Hadoop's own JobTracker, that only likes a limited set of Jobs: Map, Shuffle, Reduce.