BF-2654 was a failure that I think was due to 16 jobs all starting at exactly the same time. The tests in that particular suite tend to all follow the same setup, so all jobs simultaneously put a huge load on the disk system as they all started. To avoid this issue, I think we should put a half-second (or full-second) delay between the launching of multiple jobs in resmoke.
As we roll out more powerful instances with more CPU's, the potential for I/O overload grows.