Re: [PATCH v8 0/9] rwsem performance optimizations

[Ingo Molnar <mingo@kernel.org>]

* Tim Chen <tim.c.chen@linux.intel.com> wrote:

> Ingo,
> 
> I ran the vanilla kernel, the kernel with all rwsem patches and the 
> kernel with all patches except the optimistic spin one.  I am listing 
> two presentations of the data.  Please note that there is about 5% 
> run-run variation.
> 
> % change in performance vs vanilla kernel
> #threads	all	without optspin
> mmap only		
> 1		1.9%	1.6%
> 5		43.8%	2.6%
> 10		22.7%	-3.0%
> 20		-12.0%	-4.5%
> 40		-26.9%	-2.0%
> mmap with mutex acquisition		
> 1		-2.1%	-3.0%
> 5		-1.9%	1.0%
> 10		4.2%	12.5%
> 20		-4.1%	0.6%
> 40		-2.8%	-1.9%

Silly question: how do the two methods of starting N threads compare to 
each other? Do they have identical runtimes? I think PeterZ's point was 
that the pthread_mutex case, despite adding extra serialization, actually 
runs faster in some circumstances.

Also, mind posting the testcase? What 'work' do the threads do - clear 
some memory area? How big is the memory area?

I'd expect this to be about large enough mmap()s showing page fault 
processing to be mmap_sem bound and the serialization via pthread_mutex() 
sets up a 'train' of threads in one case, while the parallel startup would 
run into the mmap_sem in the regular case.

So I'd expect this to be a rather sensitive workload and you'd have to 
actively engineer it to hit the effect PeterZ mentioned. I could imagine 
MPI workloads to run into such patterns - but not deterministically.

Only once you've convinced yourself that you are hitting that kind of 
effect reliably on the vanilla kernel, could/should the effects of an 
improved rwsem implementation be measured.

Thanks,

	Ingo