Re: Vectored IO in XLogWrite()

On Thu, Aug 08, 2024 at 01:25:47PM -0400, Robert Haas wrote:
> I wondered whether the regression tests actually hit the iovcnt == 2
> case, and it turns out that they do, rather frequently actually.
> Making that case a FATAL causes ~260 regression test failure. However,
> on larger systems, we'll often end up with wal_segment_size=16MB and
> wal_buffers=16MB and then it seems like we don't hit the iovcnt==2
> case. Which I guess just reinforces the point that this is
> theoretically better but practically not much different.
>
> Any other votes on what to do here?

Perhaps some micro-benchmarking to prove that the patch shows benefits
when this code path is taken alone? I could fall behind that.

I'd also document the maths of the patch a bit more, it is always kind
of hard to figure out if iovec is set up right or not. And the
formulas used in the patch look magical.
--
Michael