On Sat, 2024-02-17 at 12:24 -0800, Andres Freund wrote:
> On 2024-02-17 17:48:23 +0100, Laurenz Albe wrote:
> > - Patch 0001 speeds up pq_begintypsend with a custom macro.
> > That brought the binary copy down to 3.7 seconds, which is a
> > speed gain of 15%.
>
> Nice win, but I think we can do a bit better than this. Certainly it shouldn't
> be a macro, given the multiple evaluation risks. I don't think we actually
> need to zero those bytes, in fact that seems more likely to hide bugs than
> prevent them.
>
> I have an old patch series improving performance in this area.
The multiple evaluation danger could be worked around with an automatic
variable, or the macro could just carry a warning (like appendStringInfoCharMacro).
But considering the thread in [1], I guess I'll retract that patch; I'm
sure the more invasive improvement you have in mind will do better.
> > - Patch 0001 speeds up uuid_out by avoiding the overhead of
> > a Stringinfo. This brings text mode COPY to 19.4 seconds,
> > which is speed gain of 21%.
>
> Nice!
>
> I wonder if we should move the core part for converting to hex to numutils.c,
> we already have code the for the inverse. There does seem to be further
> optimization potential in the conversion, and that seems better done somewhere
> central rather than one type's output function. OTOH, it might not be worth
> it, given the need to add the dashes.
I thought about it, but then decided not to do that.
Calling a function that converts the bytes to hex and then adding the
hyphens will slow down processing, and I think the code savings would be
minimal at best.
> > - Patch 0003 speeds up array_out a bit by avoiding some zero
> > byte writes. The measured speed gain is under 2%.
>
> Makes sense.
Thanks for your interest and approval. The attached patches are the
renamed, but unmodified patches 2 and 3 from before.
Yours,
Laurenz Albe
[1]: https://postgr.es/m/CAMkU%3D1whbRDUwa4eayD9%2B59K-coxO9senDkPRbTn3cg0pUz4AQ%40mail.gmail.com