Re: Postgres stucks in deadlock detection

On Wed, Apr 18, 2018 at 10:08 AM, Konstantin Knizhnik
<k.knizhnik@postgrespro.ru> wrote:

And it is very hard not to notice 17-times difference.
Certainly it is true in the assumption that most deadlock timeout expiration
are caused by high workload and contention, and not by real deadlocks.
But it seems to be quite common case.

If I understand this workload correctly, the contention is for the
relation extension lock.  But I think we're likely to move that out of
the heavyweight lock manager altogether in the not-too-distant future,
as proposed in https://commitfest.postgresql.org/17/1133/ ?  I'd be
interested in hearing what happens to performance with that patch
applied.

With the extension lock patch performance in increased to 1146 TPS.
So it is much better than with vanilla postgres and about 40% better than with deadlock patch (1146 vs. 719 TPS).
Profile is the following:

 33.51%  postgres                                [.] s_lock
   4.59%  postgres                                [.] LWLockWaitListLock
   3.67%  postgres                                [.] perform_spin_delay
   3.04%  [kernel]                                [k] gup_pgd_range
   2.43%  [kernel]                                [k] get_futex_key
   2.00%  [kernel]                                [k] __basepage_index
   1.20%  postgres                                [.] calculateDigestFromBuffer
   0.97%  [kernel]                                [k] update_load_avg
   0.97%  postgres                                [.] XLogInsertRecord
   0.93%  [kernel]                                [k] switch_mm_irqs_off
   0.90%  postgres                                [.] LWLockAttemptLock
   0.88%  [kernel]                                [k] _atomic_dec_and_lock
   0.84%  [kernel]                                [k] __schedule
   0.82%  postgres                                [.] ConditionVariableBroadcast
   0.75%  postgres                                [.] LWLockRelease
   0.74%  [kernel]                                [k] syscall_return_via_sysret
   0.65%  postgres                                [.] SetLatch
   0.64%  [kernel]                                [k] _raw_spin_lock_irqsave
   0.62%  [kernel]                                [k] copy_user_enhanced_fast_string
   0.59%  postgres                                [.] RelationPutHeapTuple
   0.55%  [kernel]                                [k] select_task_rq_fair
   0.54%  [kernel]                                [k] try_to_wake_up
   0.52%  [kernel]                                [k] menu_select

So definitely elimination heavy weight relation extension lock is good idea which eliminates the need for my deadlock patch ... but only in this insert test.
As I have mentioned at the beginning of this thread the same problem with deadlock detection timeout expiration we have with YSCB benchmark with zipf distribution.
Here the source of contention are tuple locks. And as far as I understand from the discussion in the mentioned thread, it is not possible to eliminate heavy weight tuple locks.

-- 
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company