On Wed, May 15, 2024 at 02:03:32PM +1200, David Rowley wrote:
> On Wed, 15 May 2024 at 13:00, Bruce Momjian <bruce@momjian.us> wrote:
> >
> > On Tue, May 14, 2024 at 03:39:26PM -0400, Melanie Plageman wrote:
> > > "Reduce system calls by automatically merging reads up to io_combine_limit"
> >
> > Uh, as I understand it, the reduced number of system calls is not the
> > value of the feature, but rather the ability to request a larger block
> > from the I/O subsystem. Without it, you have to make a request and wait
> > for each request to finish. I am open to new wording, but I am not sure
> > your new wording is accurate.
>
> I think you have the cause and effect backwards. There's no advantage
> to reading 128KB if you only need 8KB. It's the fact that doing
> *larger* reads allows *fewer* reads that allows it to be more
> efficient. There are also the efficiency gains from fadvise
> POSIX_FADV_WILLNEED. I'm unsure how to jam that into a short sentence.
> Maybe; "Optimize reading of tables by allowing pages to be prefetched
> and read in chunks up to io_combine_limit", or a bit more buzzy;
> "Optimize reading of tables by allowing pages to be prefetched and
> performing vectored reads in chunks up to io_combine_limit".
Yes, my point is that it is not the number of system calls or system
call overhead that is the advantage of this patch, but the ability to
request more of the I/O system in one call, which is not the same as
system calls.
I can use your wording, but how much prefetching to we enable now?
--
Bruce Momjian <bruce@momjian.us> https://momjian.us
EDB https://enterprisedb.com
Only you can decide what is important to you.