Re: Should io_method=worker remain the default?

On Thu, Sep 4, 2025 at 10:47 AM Jeff Davis <pgsql(at)j-davis(dot)com> wrote:
> Andres suggested that the case I brought up at the top of the thread is
> due to lock contention, so a lock free queue also sounds like a
> potential improvement. If the code is working and can be applied to
> REL_18_STABLE, then I can try it.

Not easily, unfortunately, that was a while back and I'd need to find
all the bits and rebase.

There is a preliminary step in CF5913[1]: it splits
AioWorkerSubmissionQueueLock's duties into AioWorkerControlLock for
pool management and AioWorkerSubmissionQueueLock for the queue itself.
I had a couple of different attempts at generic lock free queues, one
from a paper and one loosely based on FreeBSD's buf_ring.h, but I'm
not at all sure it's actually any good and I found it quite tricky to
integrate with the idle bitmap scheme which so far seems to be
valuable... Hmm, it might be good to try some simpler things first:
splitting AioWorkerSubmissionQueueLock into a producer lock (backends
enqueuing) and a consumer lock (workers dequeuing) must surely help a
lot. I'll try to come up with some simple patches to evaluate...

[1] https://www.postgresql.org/message-id/flat/CA%2BhUKGJLe-0E-wZ-is78CEHhjbC%3DihMVCQLoN1dmD-j05s9qRg%40mail.gmail.com#c3eb6c873b7218d5c3d5164244f7245d