Re: Move PinBuffer and UnpinBuffer to atomics - Mailing list pgsql-hackers

On Sun, Apr 10, 2016 at 8:36 AM, Alexander Korotkov <a.korotkov@postgrespro.ru> wrote:

On Sat, Apr 9, 2016 at 10:49 PM, Andres Freund <andres@anarazel.de> wrote:

On April 9, 2016 12:43:03 PM PDT, Andres Freund <andres@anarazel.de> wrote:
>On 2016-04-09 22:38:31 +0300, Alexander Korotkov wrote:
>> There are results with 5364b357 reverted.
>
>Crazy that this has such a negative impact. Amit, can you reproduce
>that? Alexander, I guess for r/w workload 5364b357 is a benefit on that
>machine as well?

How sure are you about these measurements?

I'm pretty sure.  I've retried it multiple times by hand before re-run the script.

 

Because there really shouldn't be clog lookups one a steady state is reached...

Hm... I'm also surprised. There shouldn't be clog lookups once hint bits are set.

I also tried to run perf top during pgbench and get some interesting results.

Without 5364b357:

   5,69%  postgres                 [.] GetSnapshotData

   4,47%  postgres                 [.] LWLockAttemptLock

   3,81%  postgres                 [.] _bt_compare

   3,42%  postgres                 [.] hash_search_with_hash_value

   3,08%  postgres                 [.] LWLockRelease

   2,49%  postgres                 [.] PinBuffer.isra.3

   1,58%  postgres                 [.] AllocSetAlloc

   1,17%  [kernel]                 [k] __schedule

   1,15%  postgres                 [.] PostgresMain

   1,13%  libc-2.17.so             [.] vfprintf

   1,01%  libc-2.17.so             [.] __memcpy_ssse3_back

With 5364b357:

  18,54%  postgres                 [.] GetSnapshotData

   3,45%  postgres                 [.] LWLockRelease

   3,27%  postgres                 [.] LWLockAttemptLock

   3,21%  postgres                 [.] _bt_compare

   2,93%  postgres                 [.] hash_search_with_hash_value

   2,00%  postgres                 [.] PinBuffer.isra.3

   1,32%  postgres                 [.] AllocSetAlloc

   1,10%  libc-2.17.so             [.] vfprintf

Very surprising.  It appears that after 5364b357, GetSnapshotData consumes more time.  But I can't see anything depending on clog buffers in GetSnapshotData code...