Re: Sort support for macaddr8

On Mon, Jun 3, 2019 at 1:17 PM Melanie Plageman
<melanieplageman(at)gmail(dot)com> wrote:
> I tried checking to see if there is a performance difference using the
> attached DDL based on src/test/regress/sql/macaddr8.sql. I found
> that the sort function is only exercised when creating an index (not,
> for example, when doing some type of aggregation).

As you know, it's a bit weird that we're proposing adding sort support
with abbreviated keys for a type that is 8 bytes, since you'd expect
it to also be pass-by-value on most platforms, which largely defeats
the purpose of having abbreviated keys (though sort support could
still make sense, for the same reason it makes sense to have it for
int8). However, macaddr8 isn't actually pass-by-value, and it seems
too late to do anything about that now, so abbreviated keys actually
make sense.

> With the patch applied to current master and using the DDL attached,
> the timing for creating the index hovered around 20 ms for master and
> 15 ms for the patched version.

I would expect a sufficiently large sort to execute in about half the
time with abbreviation, based on previous experience. However, I think
that this patch can be justified in a relatively straightforward way.
It extends sort support for macaddr to macaddr8, since these two types
are almost identical in every other way. We should still validate the
performance out of an abundance of caution, but I would be very
surprised if there was much difference between the macaddr and
macaddr8 cases.

In short, users should not be surprised by the big gap in performance
between macaddr and macaddr8. It's worth being consistent there.

> I think that that seems like an improvement. I was thinking of
> registering this patch for the next commitfest. Is that okay?

Definitely, yes.

> I was just wondering what the accepted way to test and share
> performance numbers is for a small patch like this. Is sharing DDL
> enough? Do I need to use pg_bench?

I've always used custom microbenchmarks for stuff like this.
Repeatedly executing a particular query and taking the median
execution time as representative seems to be the best approach.

--
Peter Geoghegan