On Fri 24-11-17 12:02:36, Byungchul Park wrote:

On Thu, Nov 16, 2017 at 02:07:46PM +0100, Michal Hocko wrote:

quoted

On Thu 16-11-17 21:48:05, Byungchul Park wrote:

quoted

On 11/16/2017 9:02 PM, Michal Hocko wrote:

quoted

for each struct page. So you are doubling the size. Who is going to
enable this config option? You are moving this to page_ext in a later
patch which is a good step but it doesn't go far enough because this
still consumes those resources. Is there any problem to make this
kernel command line controllable? Something we do for page_owner for
example?

Sure. I will add it.

quoted

Also it would be really great if you could give us some measures about
the runtime overhead. I do not expect it to be very large but this is

The major overhead would come from the amount of additional memory
consumption for 'lockdep_map's.

yes

quoted

Do you want me to measure the overhead by the additional memory
consumption?

Or do you expect another overhead?

I would be also interested how much impact this has on performance. I do
not expect it would be too large but having some numbers for cache cold
parallel kbuild or other heavy page lock workloads.

Hello Michal,

I measured 'cache cold parallel kbuild' on my qemu machine. The result
varies much so I cannot confirm, but I think there's no meaningful
difference between before and after applying crossrelease to page locks.

Actually, I expect little overhead in lock_page() and unlock_page() even
after applying crossreleas to page locks, but only expect a bit overhead
by additional memory consumption for 'lockdep_map's per page.

I run the following instructions within "QEMU x86_64 4GB memory 4 cpus":

   make clean
   echo 3 > drop_caches
   time make -j4

Maybe FS people will help you find a more representative workload. E.g.
linear cache cold file read should be good as well. Maybe there are some
tests in fstests (or how they call xfstests these days).

The results are:

   # w/o page lock tracking

   At the 1st try,
   real     5m28.105s
   user     17m52.716s
   sys      3m8.871s

   At the 2nd try,
   real     5m27.023s
   user     17m50.134s
   sys      3m9.289s

   At the 3rd try,
   real     5m22.837s
   user     17m34.514s
   sys      3m8.097s

   # w/ page lock tracking

   At the 1st try,
   real     5m18.158s
   user     17m18.200s
   sys      3m8.639s

   At the 2nd try,
   real     5m19.329s
   user     17m19.982s
   sys      3m8.345s

   At the 3rd try,
   real     5m19.626s
   user     17m21.363s
   sys      3m9.869s

I think thers's no meaningful difference on my small machine.

Yeah, this doesn't seem to indicate anything. Maybe moving the build to
shmem to rule out IO cost would tell more. But as I've said previously
page I do not really expect this would be very visible. It was more a
matter of my curiosity than an acceptance requirement. I think it is
much more important to make this runtime configurable because almost
nobody is going to enable the feature if it is only build time. The cost
is jut too high.

-- 
Michal Hocko
SUSE Labs