Re: [PATCH -mm v3 8/8] slab: do not keep free objects/slabs on dead memcg caches

[Vladimir Davydov <vdavydov@parallels.com>]

On Tue, Jun 24, 2014 at 04:38:41PM +0900, Joonsoo Kim wrote:
> On Fri, Jun 13, 2014 at 12:38:22AM +0400, Vladimir Davydov wrote:
> And, you said that this way of implementation would be slow because
> there could be many object in dead caches and this implementation
> needs node spin_lock on each object freeing. Is it no problem now?
> 
> If you have any performance data about this implementation and
> alternative one, could you share it?

I ran some tests on a 2 CPU x 6 core x 2 HT box. The kernel was compiled
with a config taken from a popular distro, so it had most of debug
options turned off.

---

TEST #1: Each logical CPU executes a task that frees 1M objects
         allocated from the same cache. All frees are node-local.

RESULTS:

objsize (bytes) | cache is dead? | objects free time (ms)
----------------+----------------+-----------------------
          64    |       -        |       373 +- 5
           -    |       +        |      1300 +- 6
                |                |
         128    |       -        |       387 +- 6
           -    |       +        |      1337 +- 6
                |                |
         256    |       -        |       484 +- 4
           -    |       +        |      1407 +- 6
                |                |
         512    |       -        |       686 +-  5
           -    |       +        |      1561 +- 18
                |                |
        1024    |       -        |      1073 +- 11
           -    |       +        |      1897 +- 12

TEST #2: Each logical CPU executes a task that removes 1M empty files
         from its own RAMFS mount. All frees are node-local.

RESULTS:

 cache is dead? | files removal time (s)
----------------+----------------------------------
      -         |       15.57 +- 0.55   (base)
      +         |       16.80 +- 0.62   (base + 8%)

---

So, according to TEST #1 the relative slowdown introduced by zapping per
cpu arrays is really dreadful - it can be up to 4x! However, the
absolute numbers aren't that huge - ~1 second for 24 million objects.
If we do something else except kfree the slowdown shouldn't be that
visible IMO.

TEST #2 is an attempt to estimate how zapping of per cpu arrays will
affect FS objects destruction, which is the most common case of dead
caches usage. To avoid disk-bound operations it uses RAMFS. From the
test results it follows that the relative slowdown of massive file
deletion is within 2 stdev, which looks decent.

Anyway, the alternative approach (reaping dead caches periodically)
won't have this kfree slowdown at all. However, periodic reaping can
become a real disaster as the system evolves and the number of dead
caches grows. Currently I don't know how we can estimate real life
effects of this. If you have any ideas, please let me know.

Thanks.