On Wed, 16 Dec 2009 19:33:15 +0900
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> wrote:
> > After we see the further works, let us discuss this patch's value.
> > 
> Ok, I'll show new version of speculative page fault.
> 

Here is a one-day patch for speculative page fault. Maybe someone can do
better than me.
  - For what direction someone fixes mmap_sem, this result seems to be a good
    for motivation.
  - Because lockless vma lookup can be a big patch, I avoid that and just uses
    caching. Then, this just helps page fault.
    In that sense, this patch is imcomplete and not very good.

Test program and unified diff is attached. After reading all e-mails,
I'd like to wait for another idea pops up. 

This test program is the best case for this patch and doesn't reflects any
real-world application behavior.

tested on x86-64, 8core, 2socket machine.

cache-miss/page fault are 32.67, 11.4, 10.6

[Before patch (2.5.32-mmotm-Dec8)
[root@bluextal memory]# /root/bin/perf stat -e page-faults,cache-misses ./multi-fault-all-split 8

 Performance counter stats for './multi-fault-all-split 8':

       15774882  page-faults
      515407936  cache-misses

   60.011299493  seconds time elapsed

[After speculative page fault]
[root@bluextal memory]# /root/bin/perf stat -e page-faults,cache-misses ./multi-fault-all-split 8

 Performance counter stats for './multi-fault-all-split 8':

       39800221  page-faults
      454331251  cache-misses

   60.002859027  seconds time elapsed

[After speculative page fault + per-thread mm_counter (already posted)]
[root@bluextal memory]# /root/bin/perf stat -e page-faults,cache-misses ./multi-fault-all-split 8

 Performance counter stats for './multi-fault-all-split 8':

       41025269  page-faults
      436365676  cache-misses

   60.007787788  seconds time elapsed


Bye.
-Kame