Re: Binary search in fmgr_isbuiltin() is a bottleneck.

On 2017-09-27 15:30:45 -0400, Robert Haas wrote:
> On Wed, Sep 27, 2017 at 1:40 PM, Andres Freund <andres(at)anarazel(dot)de> wrote:
> > I don't think you can even measure the overhead of building the
> > table. This is inserting ~8k rows in an accurately sized hashtable - a
> > vanishingly small amount of time in comparison to the backend startup
> > time (and even more so postmaster startup). My measurement shows it
> > takes about 0.4 ms to build (gdb in, query, reset oid2builtins = 0,
> > query - repeat a couple times).
>
> 0.4ms isn't negligible as a fraction of backend startup time, is it?

Well, on linux you'd only have this on postmaster startup.

> I think backend startup time is a few milliseconds.
>
> $ echo '\set x 1' > x.txt
> $ pgbench -n -C -c 1 -f x.txt -T 10
> transaction type: x.txt
> scaling factor: 1
> query mode: simple
> number of clients: 1
> number of threads: 1
> duration: 10 s
> number of transactions actually processed: 5091
> latency average = 1.965 ms
> tps = 508.866931 (including connections establishing)
> tps = 12909.303693 (excluding connections establishing)

I had tried this with an actual simplistic query, and the difference was
either nonexistant, or below in the noise. I didn't do a pgbench run
that doesn't actually do anything in the backend - doesn't seem like a
meaningful thing to measure?

Greetings,

Andres Freund