In a nutshell: * All of our stack-dependent tests inherits a common base class with functionality to create 'unique' stacks * All tests are run by a thread pool * If a test class has methods who need to be run sequentially, we define @Test(sequential=true) (default=false) * Each stack dependent test creates a unique stack by creating a default stack, then changing the mcast_addr/mcast_port (UDP) or the bind_port (TCP) * This is only done for the first channel, subsequent channels (in the same test) need to use the same config, so we create subsequent channel by passing a ref to the first channel. Subsequent channels then simply copy the config of the first channel This works very well, although there were some quirks in TestNG we had to work around. The test time for all-tests-cc in JGroups (functional tests, stack-independent and stack-dependent (4 different stacks) went down from 2.5 hours to 13 minutes ! This is still running the test suites sequentially, I expect to get under 5 minutes when switching to parallel Manik Surtani wrote: -- Bela Ban Lead JGroups / Clustering Team JBoss - a division of Red Hat