On September 26, 2018 8:53:27 PM PDT, Andrew Gierth <andrew@tao11.riddles.org.uk> wrote:
>>>>>> "Andres" == Andres Freund <andres@anarazel.de> writes:
>
> Andres> Hm, stb's results just for floating point isn't bad. The above
>Andres> numbers were for %f %f. But as the minimal usage would be about
> Andres> the internal usage of dopr(), here's comparing %.*f:
>
> Andres> snprintf time = 1324.87 ms total, 0.000264975 ms per iteration
> Andres> pg time = 1434.57 ms total, 0.000286915 ms per iteration
> Andres> stbsp time = 552.14 ms total, 0.000110428 ms per iteration
>
>Hmm. We had a case recently on IRC where the performance of float8out
>turned out to be the major bottleneck: a table of about 2.7 million
>rows
>and ~70 float columns showed an overhead of ~66 seconds for doing COPY
>as opposed to COPY BINARY (the actual problem report was that doing
>"select * from table" from R was taking a minute+ longer than expected,
>we got comparative timings for COPY just to narrow down causes).
>
>That translates to approx. 0.00035 ms overhead (i.e. time(float8out) -
>time(float8send)) per conversion (Linux server, hardware unknown).
Sounds like it could be pretty precisely be the cost measured above. My laptop's a bit faster than most server CPUs and
thetest has perfect branch prediction...
>That 66 seconds was the difference between 18s and 1m24s, so it wasn't
>a
>small factor but totally dominated the query time.
Ugh.
Andres
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.