I used Sysinternals' ProcessExplorer to check whether there are an unreasonable high number of handles or threads used, but nothing different from a "good" case. I will check out Process Monitor.

Ah, one more thing which is different between the hosts where the failure suddenly occured, and the hosts where it runs well: The "failing" hosts are virtual ones (i.e. virtual machines), while the "good" hosts are real machines.