Linus Torvalds wrote:
Well, so far I haven’t really seen any suggestions on how to improve
it much further.
3.0 will still be noticeably faster than 2.6.39 due to the other
changes made (ie the read-ahead), so yes, the regression itself is
But performance on that particular benchmark with that particular
machine is clearly not optimal, in that there are known setups that
would be faster still.
Of course, the reason for the mutex conversion was _other_ loads,
where the spinlocks had bad behavior. So it’s a balancing act. And I
suspect we’ve reached a reasonable point in that balancing, yes.
Here’s the original thread.