Re: Improve hash-agg performance

Hi,

On 2016-11-03 04:07:21 -0700, Andres Freund wrote:
> Hi,
>
> There's two things I found while working on faster expression
> evaluation, slot deforming and batched execution. As those two issues
> often seemed quite dominant cost-wise it seemed worthwhile to evaluate
> them independently.
>
> 1) We atm do one ExecProject() to compute each aggregate's
>    arguments. Turns out it's noticeably faster to compute the argument
>    for all aggregates in one go. Both because it reduces the amount of
>    function call / moves more things into a relatively tight loop, and
>    because it allows to deform all the required columns at once, rather
>    than one-by-one.  For a single aggregate it'd be faster to avoid
>    ExecProject alltogether (i.e. directly evaluate the expression as we
>    used to), but as soon you have two the new approach is faster.
>
> 2) For hash-aggs we right now we store the representative tuple using
>    the input tuple's format, with unneeded columns set to NULL. That
>    turns out to be expensive if the aggregated-on columns are not
>    leading columns, because we have to skip over a potentially large
>    number of NULLs.  The fix here is to simply use a different tuple
>    format for the hashtable.  That doesn't cause overhead, because we
>    already move columns in/out of the hashslot explicitly anyway.

I pushed a bit more polished versions of these.

Andres