The thing about large scale computational parameter tests is that they are like children. You launch them and watch for a while and they behave nicely. So you log off and go home, and then when you log back in later to check on them you find that everything went haywire while you weren't looking.
