Re: [HACKERS] Fix performance degradation of contended LWLock on NUMA

2017-11-06 18:07 GMT+03:00 Sokolov Yura <funny(dot)falcon(at)postgrespro(dot)ru>:
>
> On 2017-10-20 11:54, Sokolov Yura wrote:
>>
>> Hello,
>>
>> On 2017-10-19 19:46, Andres Freund wrote:
>>>
>>> On 2017-10-19 14:36:56 +0300, Sokolov Yura wrote:
>>>>
>>>> > > + init_local_spin_delay(&delayStatus);
>>>> >
>>>> > The way you moved this around has the disadvantage that we now do
this -
>>>> > a number of writes - even in the very common case where the lwlock
can
>>>> > be acquired directly.
>>>>
>>>> Excuse me, I don't understand fine.
>>>> Do you complain against init_local_spin_delay placed here?
>>>
>>>
>>> Yes.
>>
>>
>> I could place it before perform_spin_delay under `if (!spin_inited)` if
you
>> think it is absolutely must.
>
>
> I looked at assembly, and remembered, that last commit simplifies
> `init_local_spin_delay` to just two-three writes of zeroes (looks
> like compiler combines 2*4byte write into 1*8 write). Compared to
> code around (especially in LWLockAcquire itself), this overhead
> is negligible.
>
> Though, I found that there is benefit in calling LWLockAttemptLockOnce
> before entering loop with calls to LWLockAttemptLockOrQueue in the
> LWLockAcquire (in there is not much contention). And this way, `inline`
> decorator for LWLockAttemptLockOrQueue could be omitted. Given, clang
> doesn't want to inline this function, it could be the best way.

In attach version with LWLockAcquireOnce called before entering loop
in LWLockAcquire.

With regards,
Sokolov Yura