I've had two cases in which default parallelism has resulted in poor performance. The first was in a DBREINDEX on an index with a datetime column and the other was some performance testing of a particular query. Without going into too much detail, the end result was that both operations were faster when MAXDOP was set to 8 or less (systems were running with 16 CPUs). The DBREINDEX operation would actually peg all CPUs at 100%.