> 3 сент. 2020 г., в 23:40, Heikki Linnakangas <hlinnaka@iki.fi> написал(а):
>
> On 30/08/2020 15:04, Andrey M. Borodin wrote:
>>> 23 авг. 2020 г., в 14:39, Andrey M. Borodin <x4mmm@yandex-team.ru> написал(а):
>>>
>>> Thanks for reviewing and benchmarking, Pavel!
>> Pavel sent me few typos offlist. PFA v12 fixing these typos.
>
> In gist_indexsortbuild(), you first build all the leaf pages. Then, you read through all the index pages you just
built,to form the tuples for the next level, and repeat for all the upper levels. That seems inefficient, it would be
morebetter to form the tuples for the downlinks as you go, when you build the leaf pages in the first place. That's how
nbtsort.cworks. Also, you could WAL-log the pages as you go.
>
> In gist_indexsortbuild_flush(), can't you just memcpy() the page from
> memory to the buffer?
>
> - Heikki
Thanks for ideas, Heikki. Please see v13 with proposed changes.
But I've found out that logging page-by-page slows down GiST build by approximately 15% (when CPU constrained).
Though In think that this is IO-wise.
Thanks!
Best regards, Andrey Borodin.
