Re: computing completion tag is expensive for pgbench -S -M prepared

On 2018-06-07 17:01:47 +1200, David Rowley wrote:
> On 7 June 2018 at 16:13, Andres Freund <andres@anarazel.de> wrote:
> > in PortalRun().  That's actually fairly trivial to optimize - we don't
> > need the full blown snprintf machinery here.  A quick benchmark
> > replacing it with:
> >
> >                        memcpy(completionTag, "SELECT ", sizeof("SELECT "));
> >                        pg_lltoa(nprocessed, completionTag + 7);
> 
> I'd also noticed something similar with some recent benchmarks I was
> doing for INSERTs into partitioned tables. In my case I saw as high as
> 0.7% of the time spent building the INSERT tag. So I think it's worth
> fixing this.

I'm kinda surprised that I never noticed this until recently. I wonder
if there's a glibc or compiler regression causing this. There's some
string optimization passes, it could be that it now does less.


> I think it would be better to invent a function that accepts a
> CmdType, int64 and Oid that copies the tag into the supplied buffer,
> then make a more generic change that also replaces the code in
> ProcessQuery() which builds the tag. I'm sure there must be some way
> to get the CmdType down to the place you've patched so we can get rid
> of the if (strcmp(portal->commandTag, "SELECT") == 0) line too.

Generally sounds reasonable, I'm not sure it's worth to do the surgery
to avoid the strcmp(). That's a larger change that's somewhat
independent...


Greetings,

Andres Freund