Although only halfway on topic, since SMP related, but probably not
driver related (but might be, since disk I/O seemed involved alot), I
looked at these lately:

http://software.newsforge.com/software/04/12/27/1238216.shtml?tid=152&tid=72&tid=29
http://software.newsforge.com/software/04/12/27/1243207.shtml?tid=72&tid=29

Although this is mesured using userspace daemon MySQL (with SA threads),
although results were generally impressing for single processor systems,
several looked suboptimal when SMP was used, unfortunately.  Also when
large data was used, where I/O possibly was a bottleneck

Unfortunately I do not have profiling data, nor did the author to back
his claims and help discover the actual bottlenecks.  Since these
generated alot of I/O it possibly also could be a driver issue, but I
personally have no idea if it's that, or cost of context switching and
the big lock...  Seems that profiling such tests would be a great idea. 
The author doesn't seem closed to more tests, and the report is
recent...

Matt