On Fri, Dec 18, 2015 at 2:50 PM, Aleksander Alekseev <
a.aleks...@postgrespro.ru> wrote:
> > This idea can improve the situation with ProcLock hash table, but I
> > think IIUC what Andres is suggesting would reduce the contention
> > around dynahash freelist and can be helpful in many more situations
> > including BufMapping locks.
> I agree. But as I understand PostgreSQL community doesn't generally
> approve big changes that affects whole system. Especially if original
> problem was only in one particular place. Therefore for now I suggest
> only a small change. Naturally if it will be accepted there is no
> reason not to apply same changes for BufMapping or even dynahash itself
> with corresponding PROCLOCK hash refactoring.
> BTW could you (or anyone?) please help me find this thread regarding
> BufMapping or perhaps provide a benchmark?

You can find that in below thread:

This even contains the original idea and initial patch for replacing
spinlocks with atomic ops.  I have mentioned about the A-B-A problem
few mails down in that thread and given a link to paper suggesting how
that can be solved.  It needs more work, but doable.

> I would like to reproduce
> this issue but I can't find anything relevant in a mailing list. Also
> it seems to be a good idea to compare alternative approaches that were
> mentioned (atomics ops, group leader). Are there any discussions,
> benchmarks or patches regarding this topic?

You can find the discussion and patch related to Group leader approach
in the thread:
This patch is already committed.

> Frankly I have serious doubts regarding atomics ops since they will more
> likely create the same contention that a spinlock does. But perhaps
> there is a patch that works not the way I think it could work.

I think it is difficult to say without implementing it.  If we want
to evaluate
multiple approaches then what we can do here is I can help with writing a
patch using LWLocks and you can once evaluate the atomic ops approach
and we already have your current patch, then we can see what works out

With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com

Reply via email to