Hello Pavel,
> I checked performance - the most fast queries are execution of simple
> prepared statement
>
> prepare x as select 1;
> -- 1000000 x
> execute x;
> execute x;
> execute x;
> execute x;
>
> ## patched
> [pavel@nemesis postgresql]$ time psql -At -1 postgres -f ~/xxx.sql >
> /dev/null
>
> real 0m44,887s
> user 0m11,703s
> sys 0m6,942s
>
> This is probably the most worst case, what is possible and see some
> slowdown - in this case there is about 10% slowdown -
>
> but it is pretty untypical - the one query was less than 50 microsec. When
> there will be any IO activity or network usage, than this patch doesn't
> create any significant overhead.
Interesting. Thanks for the test.
I tried to replicate with a variant without any output: "SELECT;"
SELECT NOW() AS start \gset BEGIN; SELECT; -- 2^19 times END; SELECT NOW() - :'start';
The run time is about 19 µs per SELECT on my laptop. Over 33 runs each
alternating master with and without the patch, I got the following stats
on the measured time in seconds (average, stddev [min Q1 median Q3 max]):
- with : 9.898 ± 0.158 [9.564, 9.762, 9.936, 10.037, 10.108] - without: 9.419 ± 0.294 [8.670, 9.226, 9.533, 9.625,
9.845]
This seems consistent and significant. It suggests a 0.40-0.50 s
difference, that is about 5%, i.e. about (under) 1 µs overhead per
statement in pretty defavorable circumstances.
--
Fabien.