Re: [MM] Make mm counters per cpu instead of atomic

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Tue, 2009-11-24 at 09:17 -0600, Christoph Lameter wrote:
> On Tue, 24 Nov 2009, Zhang, Yanmin wrote:
> 
> > > True.... We need to find some alternative to per cpu data to scale mmap
> > > sem then.
> > I ran lots of benchmarks such like specjbb2005/hackbench/tbench/dbench/iozone
> > /sysbench_oltp(mysql)/aim7 against percpu tree(based on 2.6.32-rc7) on a 4*8*2 logical
> > cpu machine, and didn't find big result difference between with your patch and without
> > your patch.
> 
> This affects loads that heavily use mmap_sem. You wont find too many
> issues in tests that do not run processes with a large thread count and
> cause lots of faults or uses of get_user_pages(). The tests you list are
> not of that nature.
sysbench_oltp(mysql) is kind of such workload. Both sysbench and mysql are
multi-threaded. 2 years ago, I investigated a scalability issue of such
 workload and found mysql causes frequent down_read(mm->mmap_sem). Nick changes
it to down_read to fix it.

But this workload doesn't work well with more than 64 threads because mysql has some
unreasonable big locks in userspace (implemented as a conditional spinlock in
userspace).

Yanmin