On Tue, 2012-05-15 at 10:03 -0400, Steven Rostedt wrote:

I'll see if I can get some numbers to see how this fixes the issues with
multi threads on big boxes.

I couldn't get access to the big box, so I wrote my own test. The
attached program is what I used. It creates 400 threads and allocates a
memory range (with mmap) of 10 gigs. Then it runs all 400 threads, where
each is fighting to read this new memory. Causing lots of page faults.

I tested on a 4 CPU box with 3.4.0-rc7-rt6:

Without the patch:

map=10737418240
time = 11302617 usecs
map=10737418240
time = 11229341 usecs
map=10737418240
time = 11171463 usecs
map=10737418240
time = 11435549 usecs
map=10737418240
time = 11299086 usecs


With the patch:

map=10737418240
time = 6493796 usecs
map=10737418240
time = 6726186 usecs
map=10737418240
time = 3978194 usecs
map=10737418240
time = 6796688 usecs


So it went from roughly 11 secs to 6 secs (even had one 4sec run). This
shows that it sped up the fault access by almost half.


-- Steve