Re: PG 15 (and to a smaller degree 14) regression due to ExprEvalStep size

Hi,

On 2022-06-17 14:14:54 +1200, David Rowley wrote:
> I've put together the attached patch which removes 4 fields from the
> hashedscalararrayop portion of the struct which, once the JSON part is
> fixed, will put sizeof(ExprEvalStep) back down to 64 bytes again.

> The attached patch causes some extra pointer dereferencing to perform
> a hashed saop step, so I tested the performance on f4fb45d15 (prior to
> the JSON patch that pushed the sizeof(ExprEvalStep) up further. I
> found:

What do you think about the approach prototyped in my patch to move the hash
FunctionCallInfo into the element_tab? With a tiny bit more work that should
reduce the amount of dereferincing over the state today, while also keeping
below the limit?

> setup:
> create table a (a int);
> insert into a select x from generate_series(1000000,2000000) x;
>
> bench.sql
> select * from a where a in(1,2,3,4,5,6,7,8,9,10);
>
> f4fb45d15 + reduce_sizeof_hashedsaop_ExprEvalStep.patch
> drowley(at)amd3990x:~$ pgbench -n -f bench.sql -T 60 -M prepared postgres
> tps = 44.841851 (without initial connection time)
> tps = 44.986472 (without initial connection time)
> tps = 44.944315 (without initial connection time)
>
> f4fb45d15
> drowley(at)amd3990x:~$ pgbench -n -f bench.sql -T 60 -M prepared postgres
> tps = 44.446127 (without initial connection time)
> tps = 44.614134 (without initial connection time)
> tps = 44.895011 (without initial connection time)
>
> (Patched is ~0.61% faster here)
>
> So, there appears to be no performance regression due to the extra
> indirection. There's maybe even some gains due to the smaller step
> size.

"smaller step size"?

Greetings,

Andres Freund