Re: Sharing aggregate states between different aggregate functions

On Thu, Jul 9, 2015 at 7:44 PM, David Rowley
<david.rowley@2ndquadrant.com> wrote:
> On 15 June 2015 at 12:05, David Rowley <david.rowley@2ndquadrant.com> wrote:
>>
>>
>> This basically allows an aggregate's state to be shared between other
>> aggregate functions when both aggregate's transition functions (and a few
>> other things) match
>> There's quite a number of aggregates in our standard set which will
>> benefit from this optimisation.
>>
>
> After compiling the original patch with another compiler, I noticed a couple
> of warnings.
>
> The attached fixes these.

I did some performance tests on the patch. This patch shown good
improvement for same column aggregates. With int or bigint datatype columns,
this patch doesn't show any visible performance difference. But with numeric
datatype it shows good improvement.

select sum(x), avg(y) from test where x < $1;

Different columns:

selectivity            Head            patch
(millions)
0.1                       315              322
0.3                       367              376
0.5                       419              427
1                          551              558
2                          824              826

select sum(x), avg(x) from test where x < $1;

Same column:

selectivity            Head            patch
(millions)
0.1                       314              314
0.3                       363              343
0.5                       412              373
1                          536              440
2                          795              586

Regards,
Hari Babu
Fujitsu Australia