I've completed initial work and am attaching a spreadsheet with the results (along with the hacked pre-split build file used). The HDFS unit tests were benchmarked repeatedly and Clover was used to generate coverage reports. From these data, a subset of tests TT2 was identified that provide maximum coverage within the 10 minute constraint.
The end result is a proposed test set that runs on median in 9.22 minutes (compared to 40 minutes for all hdfs tests) and provides 88% of the original test coverage. I think this is a pretty good result. The median is used because a fair proportion of tests have shown outlier running times. Unfortunately, our initial code coverage number of 49% is lacking.
A focus was put on maintaining as much as possible the code coverage of the major HDFS classes, including Namenode, FSNamesystem, DFSClient, etc. Many of the tests within hdfs effectively act as integration tests (particularly when a test invokes the MiniDFSCluster), making it relatively easy to identify tests that stress these classes and code paths.
Another result in the spreadsheet is the identification of several tests with very large variations in run time that should be stabilized.
The build script defines another build target, run-test-ten, which runs the tests that are defined as part of the ten-minute test. You can use this to play with other combos, if you like. However, it's a pre-split build script.
Suggestions? Comments? Snark?