On Fri, Aug 15, 2025 at 3:45 PM Andres Freund <andres@anarazel.de> wrote:
> > My shared_buffers is 16GB, with pgbench scale 300.
>
> So there's actually no IO, given that a scale 300 is something like 4.7GB? In
> that case my patch could really not make a difference, neither of the changed
> branches would ever be reached?
This was an error on my part -- sorry.
I think that the problem was that I forgot that I temporarily
increased effective_io_concurrency from 100 to 1,000 while debugging
this issue. Apparently that disproportionately affected the patched
server. Could also have been an issue with a recent change of mine.
--
Peter Geoghegan
