Andi Kleen wrote:
> On Friday 24 February 2006 10:26, Nick Piggin wrote:
> 
> 
>>[aside]
>>Actually I have a scalability improvement for rwsems, that moves the
>>actual task wakeups out from underneath the rwsem spinlock in the up()
>>paths. This was useful exactly on a mixed read+write workload on mmap_sem.
>>
>>The difference was quite large for the "generic rwsem" algorithm because
>>it uses the spinlock in fastpaths a lot more than the xadd algorithm. I
>>think x86-64 uses the former, which is what I presume you're testing with?
> 
> 
> I used the generic algorithm because Andrea originally expressed some doubts 
> on the correctness of the xadd algorithms and after trying to understand them 
> myself I wasn't sure myself. Generic was the safer choice.
> 
> But if someone can show convincing numbers that XADD rwsems are faster
> for some workload we can switch. I guess they are tested well enough now
> on i386.
> 
> Or would your scalability improvement remove that difference (if it really exists)?
> 

Here is a forward port of my scalability improvement, untested.

Unfortunately I didn't include absolute performance results, but the
changelog indicates that there was some noticable delta between the
rwsem implementations (and that's what I vaguely remember).

Note: this was with volanomark, which can be quite variable at the
best of times IIRC.

Note2: this was on a 16-way NUMAQ, which had the interconnect
performance of a string between two cans compared to a modern x86-64
of smaller size, so it may be harder to notice an improvement.

-- 
SUSE Labs, Novell Inc.